Professional Documents
Culture Documents
Continous Time Stochastic Process PDF
Continous Time Stochastic Process PDF
Joseph C. Watkins
December 14, 2007
Contents
1 Basic Concepts 3
1.1 Notions of equivalence of stochastic processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Sample path properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Properties of filtrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Stopping times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Examples of stopping times . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Lévy Processes 12
3 Martingales 16
3.1 Regularity of Sample Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 Sample Path Regularity and Lévy Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3 Maximal Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.5 Law of the Iterated Logarithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4 Markov Processes 33
4.1 Definitions and Transition Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Operator Semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2.1 The Generator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2.2 The Resolvent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.3 The Hille-Yosida Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3.1 Dissipative Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3.2 Yosida Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3.3 Positive Operators and Feller Semigroups . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.4 The Maximum Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.4 Strong Markov Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.5 Connections to Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.6 Jump Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.6.1 The Structure Theorem for Pure Jump Markov Processes . . . . . . . . . . . . . . . . 53
4.6.2 Construction in the Case of Bounded Jump Rates . . . . . . . . . . . . . . . . . . . . 55
4.6.3 Birth and Death Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.6.4 Examples of Interacting Particle Systems . . . . . . . . . . . . . . . . . . . . . . . . . 59
1
CONTENTS 2
5 Stochastic Integrals 84
5.1 Quadratic Variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.2 Definition of the Stochastic Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.3 The Itô Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.4 Stochastic Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.5 Itô Diffusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
1 BASIC CONCEPTS 3
1 Basic Concepts
We now consider stochastic processes with index set Λ = [0, ∞). Thus, the process
X : [0, ∞) × Ω → S
can be considered as a random function of time via its sample paths or realizations
P {Yt = Xt } = 1.
3. Y is indistinguishable from X if
Recall that the Daniell-Kolmogorov extension theorem guarantees that the finite dimensional distribu-
tions uniquely determine a probability measure on S [0,∞) .
Exercise 1.2. If A ∈ σ{Xt ; t ≥ 0}, then there exists a countable set C ⊂ [0, ∞) so that A ∈ σ{Xt ; t ∈ C}.
RT
Thus, many sets of interest, e.g., {x; sup0≤s≤T f (xs ) > 0} or {x : 0 f (xs ) ds < ∞} are not measurable
subsets of S [0,∞) . Thus, we will need to find methods to find versions of a stochastic process so that these
sets are measurable.
and
xt− = lim xs exists for all t > 0.
s→t−
Then there exists a countable set C such that for all t ∈ [0, ∞)\C,
xt− = xt = xt+ .
Proof. Set
1
Cn = {t ∈ [0, ∞); max{d(xt− , xt ), d(xt− , xt+ ), d(xt , xt+ )} >
}
n
If Cn ∩ [0, m] is infinite, then by the Bolzano-Weierstrass theorem, it must have a limit point t ∈ [0, m]. In
this case, either xt− or xt+ would fail to exist. Now, write
∞ [
[ ∞
C= (Cn ∩ [0, m]),
m=1 n=1
Exercise 1.7. If both limits exist in the previous lemma for all t, then x− +
t = xt− for all t > 0.
Definition 1.8. Let CS [0, ∞) denote the space of continuous S-valued functions on [0, ∞). Let DS [0, ∞)
denote the space of right continuous S-valued functions having left limits on [0, ∞).
Even though we will not need to use the metric and topological issue associated with the spaces CS [0, ∞)
and DS [0, ∞), we proved a brief overview of the issues.
If we endow the space CS [0, ∞) with the supremum metric ρ∞ (x, y) = sup0≤s≤∞ d(xs , ys ), then the
metric will not in general be separable. In analogy with the use of seminorms in a Frechet to define a metric,
we set, for each t > 0,
ρt (x, y) = sup d(xmax{s,t} , ymax{s,t} ).
s
Then, ρT satisfies the symmetric and triangle inequality properties of a metric. However, if x and y
agree on [0, T ], but xs 6= ys for some s > T , then x 6= y but ρt (x, y) = 0. However, if ρt (x, y) = 0 for all t,
then x = y. Consider a bounded metric d(x ¯ 0 , y0 ) = max{d(x0 , y0 ), 1} and set ρ̄t (x, y) = sup ¯
0≤t≤t d(xs , ys ),
then we can create a metric on CS [0, ∞) which is separable by giving increasingly small importance to large
values of t. For example, we can use
Z ∞
ρ̄(x, y) = e−t ρ̄t (x, y) dt.
0
Then, (CS [0, ∞), ρ̄) is separable whenever (S, d) is separable and complete whenever (S, d) is complete.
For the space DS [0, ∞), then unless the jumps match up exactly then the distance from x to y might be
large. To match up the jumps, we introduce a continuous strictly increasing function γ : [0, ∞) → [, ∞) that
is one to one and onto and define
and Z ∞
ρ( x, y) = inf {max{ess supt | log γt0 |, e−t ργt (x, y) dt}}.
γ 0
As before, (DS [0, ∞), ρ̄) is separable whenever (S, d) is separable and complete whenever (S, d) is complete.
Exercise 1.9. CR [0, ∞) with the ρ∞ metric is not separable. Hint: Find an uncountable collection of
real-valued continuous functions so that the distance between each of them is 1.
With a stochastic process X with sample paths in DS [0, ∞), we have the following moment condition
that guarantee that X has a CS [0, ∞) version.
Proposition 1.10. If (S, d) be a separable metric space and set d1 (x, y) = min{d(x, y), 1}. Let X be a
process with sample paths in DS [0, ∞). Suppose for each T > 0, there exist α > 1, β > 0, and C > 0 such
that
E[d1 (Xt , Xs )β ] ≤ C(t − s)α 0 ≤ s ≤ t ≤ T. (1.1)
Then almost all sample paths of X belong to CS [0, ∞).
1 BASIC CONCEPTS 6
First note that by Lemma 1.5, the left side is the sum of a countable number of terms. In addition, note
that for each n ≥ 1, {t ∈ [0, T ]; d1 (Xt , Xt− ) > 1/n} is a finite set. Thus, for N sufficiently large, these points
are isolated by the 2−N partition of [0, T ]. Thus, in the limit, these jumps are captured. Consequently,
N
X 2X T
d1 (Xt , Xt− )β I{d1 (Xt ,Xt− )>1/n} ≤ lim inf d1 (Xk2−N , X(k−1)2−N )β . (1.2)
N →∞
0<t≤T k=1
Note that the left side of expression (1.2) increases as n increases. Thus, let n → ∞ to establish the
claim.
By Fatou’s lemma, and the moment inequality (1.1),
n
X 2
X T
β
E[ d1 (Xt , Xt− ) ] ≤ lim inf E[d1 (Xk2−n , X(k−1)2−n )β ]
n→∞
0<t≤T k=1
n
2
X T
≤ lim inf C2−nα
n→∞
k=1
Thus, for each T , with probability 1, X has no discontinuities on the interval [0, T ] and consequently
almost all sample paths of X belong to CS [0, ∞).
2. {Ft ; t ≥ 0} is complete if (Ω, F, P ) is complete and F0 contains all events having probability zero.
Combining the process X and the filtration {Ft ; t ≥ 0}, we have, as before:
1 BASIC CONCEPTS 7
Definition 1.13. X is adapted if Xt is Ft -measurable. In addition, a process X is {Ft }-progressive if, for
every t ≥ 0,
X|[0,t]×Ω is B[0, t] × Ft -measurable.
Exercise 1.14. 1. The filtration {Ft+ ; t ≥ 0} is right continuous.
2. If X is {Ft }-progressive, then X is {Ft }-adapted and measurable. Hint. Approximate X on [0, t] × Ω
by Xsn (ω) = Xmin{t,([ns]+1)/n)} (ω).
3. If X is {Ft }-adapted and right continuous, then is {Ft }-progressive.
4. Let X be {Ft }-progressive and let f : S → R be bounded and measurable. Show that f ◦ X is {Ft }-
Rt
progressive and At = 0 f (Xs ) ds is {Ft }-adapted.
(s, ω) 7→ Xs (ω)
Lemma 1.18. A [0, ∞]-valued random variable τ is an Ft+ -stopping time if and only if {t < τ } ∈ Ft for
every t ≥ 0.
Proof. (necessity)
∞
[ 1
{τ < t} = {τ ≤ t − } ∈ Ft ⊂ Ft+ .
n=1
n
(sufficiency) For n ≥ m,
1
{τ < t + } ∈ Ft+1/m .
n
Therefore,
∞ ∞
\ 1 \
{τ ≤ t} = {τ < t + }∈ Ft+1/m ⊂ Ft+ .
n=1
n m=1
{τ = t} ∈ Ft .
lim sup τn
n→∞
Now use
lim inf τn = sup inf τn and lim sup τn = inf sup τn .
n→∞ m n≥m n→∞ m n≥m
1 BASIC CONCEPTS 9
Proposition 1.21. Every Ft -stopping time is the nonincreasing limit of bounded stopping times.
Proof. Take the stopping times τn = min{τ, n} for n = 1, 2, . . . .
Definition 1.22. Call a stopping time discrete if its range is almost surely countable.
Proposition 1.23. Every Ft -stopping time is the nonincreasing limit of discrete stopping times.
Proof. For n = 1, 2, . . ., choose the nested collection of sequences 0 = tn0 < tn1 < · · · with limk→∞ tnk = ∞,
and limn→∞ supk (tnk+1 − tnk ) = 0. For an Ft+ -stopping time τ , define
n
tk if tnk−1 ≤ τ < tnk
τn =
∞ if τ = ∞.
Then clearly τn is discrete and τ = limn→∞ τn . To see that τn is a stopping time, set γn (t) = max{tnk ; tnk ≤
t}. Then
{τn ≤ t} = {τn ≤ γn (t)} = {τ ≤ γn (t)} ∈ Fγn (t) ⊂ Ft .
Exercise 1.24. Let {τn ; n ≥ 1} be a sequence of Ft -stopping times and let τ = limn→∞ τn .
1. If {τn ; n ≥ 1} is nondecreasing, then τ is an Ft -stopping time.
2. If {τn ; n ≥ 1} is nonincreasing, then τ is an Ft+ -stopping time.
Definition 1.25. For a stochastic process X and a finite stopping time τ Define the stopped process
Xtτ = Xmin{τ,t}
Now, define
Cs,t = {C ∩ ([0, t] × {min{τ, t} ≥ s}); C ∈ B[0, t] × Fs }.
1 BASIC CONCEPTS 10
Thus, B × A ∈ Cs,t .
Now, write
{(s, ω) ∈ [0, t] × Ω; Xmin{τ (ω),s} (ω) ∈ Γ}
= {(s, ω) ∈ [0, t] × Ω; Xmin{τ (ω),s} (ω) ∈ Γ, min{τ (ω), t} ≤ s ≤ t} ∪ {(s, ω); Xs (ω) ∈ Γ, s < min{τ, t}}.
To see that the first term is in B[0, t] × Gt , write it as
(u, ω) → (τ (ω) + u, ω)
(u, ω) → Xτ +u (ω)
The first exit from B after time s is the first entrance to B c . If s = 0, then it is suppressed in the
notation.
Proposition 1.28. Suppose that X is a right continuous Ft -adapted process and that σ is an Ft -stopping
time.
1. If A is closed and X has left limits at each t > 0 or if A is compact then τc (A, σ) is an Ft -stopping
time.
2. If A is open then τe (A, σ) is an Ft+ -stopping time.
Proof. If A is open, we have by the right continuuity of X,
[
{τe (A, σ) < t} = {Xs ∈ A} ∩ {σ < s} ∈ Ft
s∈Q∩[0,t]
proving 2.
Under the conditions of part 1,
τc (A, σ) = lim τe (A1/n , σ)
n→∞
proving 1.
2 LÉVY PROCESSES 12
2 Lévy Processes
We begin our study of continuous time stochastic process with the continuous time analog of random walks.
Definition 2.1. An Rd -valued process X is call a Lévy process or a process with stationary independent
increments if
1. (stationary increments) The distribution of Xt+s − Xt does not depend on t.
2. (independent increments) Let 0 = t0 < t1 < · · · < tn . Then the random variables
{Xtj − Xtj−1 ; j = 1, . . . , n}
are independent.
3. (stochastic continuity)
Xt − X0 →P 0 as t → 0.
Exercise 2.2. 1. Any linear combination of independent Lévy processes is a Lévy process.
2. If X has finite mean, X0 = 0 and µ = EX1 , then EXt = µt.
3. If, in addition, X has finite variance and σ 2 = Var(X1 ), then Var(Xt ) = σ 2 t.
Definition 2.3. A Poisson process N with parameter λ is a Lévy process with N0 = 0 and
(λt)n −λt
P {Nt = n} = e .
n!
Exercise 2.4. Show that the definition of a Poisson process satisfies the compatibility condition in the
Daniell-Kolmogorov extension theorem.
Definition 2.5. Let {Yk , k ≥ 1} be independent and identically distributed Rd -valued random variables and
let N be a Poisson process, then
XNt
Xt = Yk
k=1
where Im is the modified Bessel function of order m. Hint. The modified Bessel functions of integer order
Im (x) have generating function exp(x(z + z −1 )/2).
2 LÉVY PROCESSES 13
where ν and φY are, respectively, the distribution and the characteristic function for the Yk .
Exercise 2.8. For a Lévy process X let νs and φs (u) denote, respectively, the distribution and the charac-
teristic function for the Xs . Then, φs+t = φs φt and consequently νs+t = νs ∗ νt .
Use the dominated convergence theorem to show that the stochastic continuity of a Lévy process implies
that φs is right continuous and so, by the functional relationship in the exercise above,
for some ψ.
Definition 2.9. The function ψ is called the characteristic exponent for the Lévy process.
Definition 2.10. A one dimensional Brownian motion B with parameters µ and σ 2 is a Lévy process with
B0 = 0 and
(x − µt)2
Z
1
P {Bt ∈ A} = √ exp − dx.
σ 2πt A 2σ 2 t
If µ = 0 and σ 2 = 1, the B is called a standard Brownian motion.
For d-dimensional Brownian motion B, we require µ ∈ Rd and a symmetric, positive semi-definite matrix
Σ, Then, B is a Lévy process and if B0 = 0, then
Exercise 2.11. 1. Show that Brownian motion as defined above satisfies the consistency condition in the
extension theorem.
2. Find the characteristic exponent for Brownian motion.
Exercise 2.12. Let B be a d-dimensional Brownian motion. Show that there exists an invertible matrix C ,
a vector µ, and a d-dimensional process B̃t = (B̃t1 , . . . , B̃td ) in which the component are independent standard
Brownian motions so that Bt = C B̃t + µt.
As a consequence, we sometimes call B a Brownian motion if Bt = C B̃t for some d × d-matrix C.
Exercise 2.13. Let B be a standard Brownian motion. Then the following are also standard Brownian
motions.
2 LÉVY PROCESSES 14
1. −B,
2. {Bt+t0 − Bt0 ; t ≥ 0} for t0 > 0.
3. {aBt/a2 ; t ≥ 0} for a > 0.
Proof. Set B ∗ = sup{Bt ; t ∈ Q ∩ [0, ∞)}. Then, for any a ∈ Q+ , B ∗ and aB ∗ have the same distribution.
Consequently, B ∗ is concentrated on the set {0, ∞}. Set p∗ = P {B ∗ = 0}. Then
Thus, for any n, Xt can be written as the sum of n independent identically distributed random variables.
This motivates the following:
Definition 2.15. A random variable Y is said to have an infinitely divisible distribution if for every n, there
exists n independent identically distributed random variables, Y1,n , . . . , Yn,n such that
Y = Y1,n + · · · + Yn,n .
The Lévy-Khintchine formula gives the characteristic function for any infinitely divisible distribution on
Rd , Z
1 ihu, yi
exp ψ(u) = exp ihβ, ui − hu, Σui + eihu,yi − 1 − ν(dy) .
2 1 + |y|2
For this formula, there is some flexibility in the last function χ(y) = y/(1 + |y|2 ). We need χ to be anti-
symmetric, bounded and have derivative 1 at the origin. We can choose β to be any d-dimensional vector
and Σ to be any symmetric non-negative definite d × d-matrix. The measure ν, called the Lévy measure can
be any Radon measure satisfying
1. ν{0} = 0, and
R |y|2
2. 1+|y| 2 ν(dy) < ∞.
2 LÉVY PROCESSES 15
Now, clearly exp(ψ(u)/n) is the characteristic function of a random variable. The sum of n such inde-
pendent random varibles has characteristic function exp ψ(u). Thus, we have the converse that any infinitely
divisible
R distribution can be realized by X1 for some Lévy process X.
|y|
If 1+|y| 2 ν(dy) < ∞, then we can write
Z
1
exp ψ(u) = exp(ihβ̃, ui − hu, Σui + (eihu,yi − 1) ν(dy)).
2
where Z
y
β̃ = β + ν(dy)
1 + y2
This is the characteric function of a random variable that is the independent sum of a normal random
variable and a compound Poisson process evaluated at time 1.
If λ = ν(Rd \{0}) < ∞, and if CC T = Σ, then the process
Nt
X
Xt = βt + C B̃t + Yn − µλt
n=1
the X is a Lévy process with characteristic exponent ψ where B̃t = (B̃t1 , . . . , B̃td ) with components that are
independent standard Brownian motions, N is a Poisson process with parameter λ and {Yn ; n ≥ 1} is an
independent sequence of random variables with common distrobution ν/λ.
3 MARTINGALES 16
3 Martingales
The definition of martingale in continuous time is exactly the analog of the discete time definition.
Definition 3.1. A real valued process X with E|Xt | < ∞ for all t ≥ 0 and adapted to a filtration {Ft ; t ≥ 0}
is an Ft -martingale if
E[Xt |Fs ] = Xs for all t > s,
an Ft -submartingale if
E[Xt |Fs ] ≥ Xs for all t > s,
and an Ft -;gale if the inequality above is reversed.
Remark 3.2. Many previous facts about discrete martingales continue to hold for continuous time martin-
gales.
1. X is an Ft martingale if and only if for every s < t, E[Xt ; A] = E[Xs ; A], for all A ∈ Fs .
2. Suppose X is an Ft -martingale, φ is convex and E|φ(Xt )| < ∞ for all t ≥ 1. Then φ ◦ X is an
Ft -submartingale.
3. Suppose X is an Ft -submartingale, φ is convex and non-decreasing, E|φ(Xt )| < ∞ for all t ≥ 0. Then
φ ◦ X is an Ft -submartingale.
Exercise 3.3. An adapted stochastic process M is a martingale if and only if
EMτ = EM0
Proposition 3.6. Let τ and σ be two Ft -stopping times taking values in a finite set F = {t1 < · · · < tm }.
If X is an Ft -submartingale, then
E[Xτ |Fσ ] ≥ Xmin{τ,σ} .
Proof. We show that for every A ∈ Fσ ,
E[Xmin{τ,tk+1 } |Ftk ] = E[Xmin{τ,tk+1 } I{τ >tk } |Ftk ] + E[Xmin{τ,tk+1 } I{τ ≤tk } |Ftk ]
= E[Xtk+1 |Ftk ]I{τ >tk } + Xmin{τ,tk } I{τ ≤tk }
≥ Xtk I{τ >tk } + Xmin{τ,tk } I{τ ≤tk } = Xmin{τ,tk }
Thus, E[Xτ |Fti ] = E[Xmin{τ,tm } |Fti ] ≥ Xmin{τ,tm−1 } . Use this for the frist step and the tower property
of σ-algebras in the induction step to complete the proof.
Example 3.7. Let N be a Poisson process with parameter λ. Let τn = inf{t ≥ 0; Nt = n}.
1. Fix a time t, then min{τn , t}is bounded, so by the optional sampling theorem,
Now, use the monotone convergence theorem on each of these terms and the fact that τn < ∞ almost
surely to obtain
n
0 = ENτn − λEτn and, therefore, Eτn =
λ
2. Consider
exp(iuNt − λt(eiu − 1)),
the exponential martingale for the Poisson process. Then
and thus
e−iun = E[exp(−λτn (eiu − 1)].
Set
v+λ
v = λ(eiu − 1) or u = −i log( )
λ
3 MARTINGALES 18
yielding
λ n
Ee−vτ = ()
λ+v
This is the n-th power of the Laplace transform of an exponential random variable and, hence, the
Laplace transform of a Γ(n, λ) random variable.
Exercise 3.8. Using N and τn as in the example above. Show that τn+1 − τn is an exponential random
variable independent of FτNn t see that {τn+1 − τn ; n ≥ 1} is an independent and identically distributed
sequence of exponential random variables.
Exercise 3.9. Note that {Nt < n} = {τn > t}. Use this to obtain the density of a Γ(n, λ) random variable.
τ = min{u ∈ F ; Xu ≤ −x}.
By the proposition above, E[Xt |Fτ ] ≥ Xmin{τ,t} . In addition, if τ < ∞, then min τ, t} = τ .
Now simplify.
Corollary 3.11. Let X be a submartingale, t > 0 and C a countable subset of [0, t]. Then for each x > 0,
1
P {max Xs ≥ x} ≤ EXt+ ,
s∈C x
and
1
P {min Xs ≤ −x} ≤ (EXt+ − EX0 ).
s∈C x
Proof. Choose finite subests F1 ⊂ F2 ⊂ · · · so that C = ∪∞
n=1 Fn . Then, for 0 < x̃ < x,
1
P {max Xs ≥ x} ≤ lim P {max Xs ≥ x̃} ≤ EXt+ .
s∈C n→∞ s∈Fn x̃
Now, let x̃ → x.
3 MARTINGALES 19
We can now make similar analogies to the discrete time case for upcrossings. For a stochastic process X,
define U (a, b, F ) to be the number of upcrossings of the interval (a, b) by X restricted to a finite set F . For
C countable, as before, write C = ∪∞ n=1 Fn and define Then U (a, b, Fn ) is a monotone increasing sequence.
Call its limit U (a, b, C).
Exercise 3.13. Show that the definition above is independent of the choice of the finite sets Fn .
Theorem 3.14 (Doob’s upcrossing inequality). Let X be a submartingale and let t > 0. For a countable
subset C ⊂ [0, t],
E(Xt − a)+
EU (a, b, C) ≤ .
b−a
Proof. Choose Fn as above, then by the discrete time version of the uncrossing inequality
E(Xt − a)+
EU (a, b, Fn ) ≤ .
b−a
Now, use the monotone convergence theorem.
Corollary 3.15. Let D be a countable dense subset of [0, ∞). Then
P {U (a, b, D ∩ [0, t]) < ∞} = 1.
Remark 3.16. Set
∞
\ \
Ω0 = { sup Xt < ∞} ∩ { inf Xt ≥ −∞} ∩ {U (a, b, D ∩ [0, n]) < ∞} .
s∈D∩[0,n] s∈D∩[0,n]
n=1 a<b,a,b∈D
Then, P (Ω0 ) = 1.
Exercise 3.17. For ω ∈ Ω0 ,
Xt+ (ω) = lim Xs (ω)
s→t+,s∈D
Proposition 3.18. Let X be a submartingale and define X + as above. Then Γ = {t ≥ 0; P {Xt+ 6= Xt−
+
}>
+
0} is countable. P {Xt = Xt } = 1 for t ∈
/ Γ and
Xt (ω) if t ∈ Γ
X̃t (ω) =
Xt+ (ω) if t ∈
/Γ
defines a modification of X almost all of whose sample paths have right and left limits at all t ≥ 0 and are
right continuous at all t ∈
/ Γ.
3 MARTINGALES 20
Proof. Consider the space L∞ (Ω) of real-valued random variables on Ω and define the metric
Now X + has right and left hand limits in this metric and so by the lemma Γ is countable.
Choose a real number a, then φa (x) = max{x, a} is an increasing convex function and so φa (Xt ) is a
submartingale. Consequently,
we see that {φa (Xs ); 0 ≤ s ≤ t} is also uniformly integrable. Therefore, almost sure convergence implies
convergence in L1 and
+
/ Γ, Xs+ = Xs−
For s ∈ a.s and
Thus, the non-negative random variable E[φa (Xs+ )|FsX ] − φa (Xs ) has zero expectation and is consequently
0 almost surely.
+ +
Using again that Xs+ = Xs− a.s. and that Xs− is FsX -measurable,
Thus, for almost all ω, if Xs (ω) > a, then Xs (ω) = Xs+ (ω). Because this holds for all a ∈ R,
Thus,
Xt+ (ω) = lim X̃s (ω)
s→t+
Remark 3.19. In most of the situations that we will encounter, Γ = ∅ and if the process is defined, as in
the case of Lévy processes, by its finite dimensional distributions, then we can take the version with sample
paths in DS [0, ∞).
Exercise 3.20. The convergence under the metric ρ is equivalent to convergence in probability.
Corollary 3.21. Let Z ∈ L1 , then for any filtration {Ft ; t ≥ 0}, and for all t ≥ 0,
1
E[Z|Fs ] →L E[Z|Ft+ ] as s → t + .
Proof. Let Xt = E[Z|Ft+ ]. Then X is a martingale. By the proposition above, X has right limits a.s. at
each t ≥ 0. X is uniformly integrable. Consequenty,
Note that Xt+ is Ft+ -measurable. Take A ∈ Ft+ , then for s > t, Ft+ ⊂ Fs and E[Xs ; A] = E[Z; A].
Consequently,
E[Xt+ ; A] = lim E[Xs ; A] = E[Z; A].
s→t+
t → Xt+
lim E[|Xs − Xt |] = 0.
s→t+
Proof.
Theorem 3.23 (Optional sampling formula). Let X be a right continuous Ft -submartingale and let τ and
σ be Ft -stopping tmes. Then for each t > 0,
Proof. We have the theorem in the case that τ and σ are discrete stopping times. Now, let {σn , τn ; n ≥ 1}
be a sequence of nonincreasing discrete stopping times that, converge, respectively to τ and σ.
As before for a, set φa (x) = max{x, a} and note thatφa (Xt ) is a submartingale. Then,
Use the fact that Fσ ⊂ Fσn and the tower property to conclude that
Note that
a ≤ φa (Xmin{τn ,t} ) ≤ E[φa (Xt )|Fτn ] and a ≤ φa (Xmin{σn ,τn ,t} ) ≤ E[φa (Xt )|Fmin{σn ,τn } ].
Consequently, as before, {φa (Xmin{τn ,t} ); n ≥ 1} and {φa (Xmin{σn ,τn ,t} )} are uniformly integrable. Now
let n → ∞ and use the right continuity of X to obtain
Now let, a → −∞. The second part follows are in the discrete time case.
We have similar theorems to the discrete time case on convergence as t → ∞.
then
lim Xt = X∞
t→∞
and hence
1
E[Xt |Fτn + ]I{τn ≤t} , ≤ ,
n
and, taking n → ∞, we have upon taking expected values that
Now, let n → ∞. For the third term, use the bounded convergence theorem to obtain
0 = bP {Bτ = b} − aP {Bτ = a}
or
a
Eτ = P {Bτ = b} = .
b+a
Now, use the martingale Bt2 − t to obtain
2
0 = EBmin{τ,n} − E min{τ, n}.
For the first term use the bounded convergence theorem and for the second use the monotone convergence
theorem as n → ∞ to see that
a b
EBτ2 = b2 + a2 = ab.
b+a b+a
For the second term use the monotone convergence theorem as n → ∞ to see that
Eτ = ab.
τx = inf{t ≥ 0; Xt = x}
Now
exp(uXt − αt) = exp(uBt − (α − uµ)t)
3 MARTINGALES 24
In addition, the Laplace transform can be inverted to see that τx has density
x (x − µt)2
fτx (t) = √ exp(− ).
2πt3 2t
Exercise 3.30. Let B be standard Brownian motion, a > 0, and τ = inf{t ≥ 0; |Bt | = a}.
√
1. E[exp(−ατ )] = sech(a 2α).
2. Eτ 2 = 5a4 /3. Hint: Prove that Bt4 − 6tBt2 + 3t2 is a martingale.
Proposition 3.31. Let X be a right continuous Lévy process. Then for each s ≥ 0, the process X̃t =
X
Xs+t − Xs is a Lévy process with the same finite dimensional distribution as X that is independent of Fs+ .
X X X X
Proof. For each > 0, Fs+ ⊂ Fs+ , Xs+t+ − Xs+ is a Lévy process independent of Fs+ . Let A ∈ Fs+
Choose a right continuous modification of X, times 0 = t0 < t1 < · · · < tn , and bounded continuous functions
fj .
n
Y n
Y
E[ fj (Xs+tj − Xs+tj−1 ); A] = lim E[ f (Xs+tj + − Xs+tj−1 + ); A]
→0+
j=1 j=1
Yn
= lim E[ f (Xs+tj + − Xs+tj−1 + )]P (A)
→0+
j=1
Yn
= lim E[ f (Xtj − Xtj−1 )]P (A)
→0+
j=1
n
Y
= E[ f (Xtj − Xtj−1 )]P (A)
j=1
3 MARTINGALES 25
Proposition 3.32. Let X be a right continuous Lévy process and τ an almost surely finite stopping time.
Then the process X̃s = Xs+τ − Xτ is a Lévy process with respect to the filtration {FτX+s ; s ≥ 0} having the
same finite dimensional distributions as Xs − X0 and is independent of FτX+ .
Proof. Consider times 0 = t0 < t1 < · · · < tn , bounded continuous functions f1 , f2 , . . . , fn , and A ∈ FτX+ .
For the case that τ is a discrete stopping time with range {sk ; k ≥ 1}, we have
n
Y ∞
X n
Y
E[ fj (Xτ +tj − Xτ +tj−1 ); A] = E[ fj (Xτ +tj − Xτ +tj−1 ); A ∩ {τ = sk }]
j=1 k=1 j=1
X∞ Yn
= E[ fj (Xsk +tj − Xsk +tj−1 ); A ∩ {τ = sk }]
k=1 j=1
X∞ Yn
= E[ fj (Xtj − Xtj−1 )]P (A ∩ {τ = sk })
k=1 j=1
n
Y
= E[ fj (Xtj − Xtj−1 )]P (A)
j=1
Note that A ∩ {τ = sk } ∈ Fsk + and thus the third line follows from the second by the previous proposition.
For the general case, let {τm ; m ≥ 1} be a decreasing sequence of stopping times with limit τ . Pick
A ∈ Fτ ⊂ Fτm . Now, apply the identity above to the τm , use right continuity and the bounded convergence
theorem to obtain the result.
X
Corollary 3.33 (Blumenthal 0-1 law). For a Lévy process X, with X0 = 0 every event in F0+ has probability
0 or 1.
X X
Proof. Fix A ∈ F0+ , then, in addition, A ∈ σ{Xt ; t ≥ 0} = σ{Xt − X0 ; t ≥ 0}. Then F0+ and σ{Xt − X0 ; t ≥
0} are independent σ-algebras and thus P (A) = P (A ∩ A) = P (A)P (A) and so P (A) = 0 or 1.
X
In particular, if τ is an Ft+ -stopping time then
P {τ = 0} = 0 or 1.
Example 3.34. 1. Suppose that f (t) > 0 for all t > 0. Then
Bt
lim sup = c a.s.
t→0+ f (t)
√
for some c ∈ [0, ∞]. If f (t) = t, the Bt /f (t) is a standard normal and so the limit above cannot be
finite. Consequently, c = ∞. If f (t) = tα , α > 1/2, then Bt /f (t) is normal, mean 0, variance t1−2α
and so the limit above cannot be positive. Consequently, c = 0.
2. Let τ = inf{t > 0; Bt > 0}. Then P {τ ≤ t} ≥ P {Bt > 0} = 1/2. Thus,
P {τ = 0} = lim P {τ ≤ t} =
6 0,
t→0+
thus it must be 1.
3 MARTINGALES 26
Theorem 3.35 (Reflection principle). Let a ∈ R and let B be a continuous standard Brownian motion.
Define the stopping time τa = inf{t > 0; Bt > a}. The process
Bt if t < τa ,
B̃t =
2a − Bt if t ≥ τa .
Proof.
3. Show that {τa ; a ≥ 0} is a Lévy process that does not have a finite stopping time.
3 MARTINGALES 27
Proof. The first statement follows form the corresponding result for a countable collection of times in [0, t]
and right continuity.
For the second statement. set τ = inf{t ≥ 0; Xs > x}. Then τ is an Ft+ -stopping time. The right
continuity of X implies that Xτ ≥ x whenever τ < ∞. Consequently,
Moreover, these events have equal probability for all but countably many x > 0 and hence
This subsequential limit along the integers is 0 by the strong law of large numbers. The following claim
proves the theorem.
3 MARTINGALES 28
Claim.
maxn≤t≤n+1 |Bt − Bn |
lim = 0 a.s.
n→∞ n
Exercise 3.41. Let Φ be the cumulative distribution function for the standard normal. Then, for a > 0
1 1
Φ(−a) = 1 − Φ(a) ≥ √ exp(− a2 )a−1 (1 − a−2 ).
2π 2
Hint. Use exp(−z 2 /2) > (1 − 3z −4 ) exp(−z 2 /2) and find the antiderivative.
3.4 Localization
To extend our results to a broader class of processes, we introduce the concept of localization.
This next proposition gives us our first glimpse into the stochastic calculus.
Proposition 3.44. Suppose that Y is a right continuous Ft -local martingale, and that V is a real-valued
Ft -adapted, bounded continuous process having locally bounded variation. Then
Z t
Mt = Vt Yt − Ys dVs
0
is an Ft -local martingale.
Remark 3.45. This uses the fact that we will later learn that integrals of martingales are martingales.
Then, M can be found from an integration by parts formula
Z t Z t
Vs dVs = Vt Yt − V0 Y0 − Yu dVu .
0 0
Exercise 3.46 (summation by parts). Let {ak ; k ≥ 0} and {bk ; k ≥ 0} be two real valued sequences, then
n−1
X n−1
X
ak (bk+1 − bk ) = an bn − am bm − bk+1 (ak+1 − ak ).
k=m k=m
Proof. (of the proposition) Let {τn ; n ≥ 1} be a localizing sequence for Y so that τn is less than the contact
time with the set (−n, n)c .
Write |V |t for the variation of V on the interval [0, t]. Because V is continuous, σn , the first contact time
of |V | into [n, ∞) is a stopping time. Because |V | is continuous, σn → ∞ as n → ∞.
Write γn = min{σn , τn }, we show that M γn is a martingale, i. e,
" #
Z min{t+s,γn }
γn γn
E Vt+s Yt+s − Ysγn Vsγn − Yuγn dVuγn |Ft = 0. (3.1)
min{t,γn }
Let t = u0 < u1 < · · · < um = t + s be a partition of the interval [t, t + s]. Then, by the tower property,
and the fact that V is adapted and thatYγn is a martingale, we see that
m−1
X m−1
X
E[ Vuγkn (Yuγk+1
n
− Yuγkn )|Ft ] = E[Vuγkn E[(Yuγk+1
n
− Yuγkn )|Fuk |Ft ] = 0.
k=0 k=0
The sum in the conditional expectation converges almost surely to the Riemann-Stieltjes integral
Z t+s
Yuγn dVuγn .
t
3 MARTINGALES 30
To complete the proof, we must show that the process is uniformly integrable. However,
n−1
X m−1
X γn γn γn γn
Yuγkn (Vuγk+1
n
− Vuγkn ) ≤ |Yuk ||Vuk+1 − Vuγkn | ≤ max{n, Yt+s }(|Vγn |t+s − |V γn |t ) ≤ max{n, Yt+s }n
k=0 k=0
which is integrable.
Thus, the Riemann sums converge in L1 and we have both that (3.1) holds and that Mγn is uniformly
integrable.
Consider the exponential martingale Zt (α) = exp(αBt − 21 α2 t). Then by Doob’s submartingale maximal
inequality,
1
P { sup (Bs − αs) > β} = P { sup Zs (α) > eαβ } ≤ e−αβ EZt (α) = e−αβ .
0≤s≤t 2 0≤s≤t
Fix the θ and δ in (0, 1) and apply the inequality above to the sequences
1
tn = θn , αn = θ−n (1 + δ)h(θn ), βn = h(θn ).
2
Then,
1 −n 1
αn βn = θ (1 + δ)h(θn )2 = (1 + δ) log log θ−n = (1 + δ)(log n + log log ).
2 θ
Consequently, for γ = (log(1/θ))(1+δ) ,
1
P { sup (Bs − αn s) > βn } ≤ γn−(1+δ) .
0≤s≤tn 2
Because δ > 0, these probabilities have a finite sum over n. Thus, by the first Borel-Cantelli lemma
1
P { sup (Bs − αn s) > βn i.o.} = 0.
0≤s≤tn 2
Thus, we have, for ω in an event having probability 1, there exists N (ω) so that
1
sup (Bs − αn s) > βn
0≤s≤tn 2
3 MARTINGALES 31
for all n ≥ N (ω). For such n, consider t in the interval (θn+1 , θn ]. Then
1 1 1
Bt (ω) ≤ sup Bs (ω) ≤ βn + αn θn = (2 + δ)h(θn ) ≤ θ−1/2 (2 + δ)h(t).
0≤s≤θ n 2 2 2
Consequently,
Bt (ω) 1
lim sup ≤ θ−1/2 (2 + δ).
t→0+ h(t) 2
Now, use the fact inf θ,δ∈(0,1) 21 θ−1/2 (2 + δ) = 1 to obtain
Bt (ω)
lim sup ≤ 1.
t→0+ h(t)
Part 2.
Bt
P lim sup ≥ 1 = 1.
t→0+ h(t)
4 Markov Processes
4.1 Definitions and Transition Functions
Definition 4.1. A stochastic process X with values in a metric space S is called a Markov process provided
that
P {Xs+t ∈ B|FtX } = P {Xs+t ∈ B|Xt } (4.1)
for every s, t ≥ 0 and B ∈ B(S).
If {Ft ; t ≥ 0} is a filtration and FtX ⊂ Ft , then we call X a Markov process with respect to the filtration
{Ft ; t ≥ 0} if the (4.1) holds with FtX replaced by Ft .
The probability measure α given by α(B) = P {X0 ∈ B} is called the initial distribution of X.
Exercise 4.2. 1. If X is Markov process with respect to the filtration {Ft ; t ≥ 0} then it is a Markov
process.
2. Use the standard machine to show that the definition above is equivalent to
The transition function P is a transition function for a time-homogeneous Markov process X, if, for
every s, t ≥ 0 and B ∈ B(S),
P {Xt+s ∈ B|FtX ] = p(s, Xt , B).
Exercise 4.4. Show, using the standard machine, the the identity above is equivalent to
Z
X
E[f (Xt+s |Ft ] = f (y)p(s, Xt , dy)
S
To see how the Chapman-Kolmgorov equation arises, note that for all s, t, u ≥ 0 and B ∈ B(S),
p(t + s, Xu , B) = P {Xu+t+s ∈ B|FuX } = E[P {Xu+t+s ∈ B|Fu+tX
}|FuX ]
Z
X
= E[p(s, Xu+t , B)|Fu ] = p(s, y, B)p(t, Xu , dy).
S
Exercise 4.5. Let X be a Lévy process with νt (B) = P {Xt ∈ B}. Then the transition function
p(t, x, B) = νt (B − x).
In this case, the Chapman-Kolmogorov equations state that νs+t = νs ∗ νt .
The transition function and the initial distribution determine the finite dimensional distributions of X
by
P {X0 ∈ B0 , Xt1 ∈ B1 , . . . , Xtn ∈ Bn }
Z Z Z
= ··· p(tn − tn−1 , xn−1 , Bn )p(tn−1 − tn−2 , xn−2 , dxn−1 ) · · · p(t1 , x0 , dx1 )α(dx0 ).
B0 B1 Bn−1
Exercise 4.6. Show that the consistency condition in the Daniell-Kolmogorov are satisfied via the Chapman-
Kolmogorov equation.
This gives us a unique measure on (S [0,∞) , B(S [0,∞) ). Denote probabilities with respect to this measure
Pα and Px for Pδx .
Rarely are we able to write this transition function explicitly. In two well known cases, Brownian motion
and the Poisson process we can.
Example 4.7. For Brownian motion,
(y − x)2
Z
1
P (t, x, B) = √ exp − dt, t ≥ 0, x ∈ R, B ∈ (R).
2πt B 2t
For the Poisson process with parameter λ,
(λt)y−x −λt
P (t, x, {y}) = e , t ≥ 0, y ≥ x ≥ 0, x, y ∈ N.
(y − x)!
The kernels P (s, ·, ·) are naturally associated with the operator on the bounded measurable functions.
Definition 4.8. The transition operator for the transition function P is defined by
Z
T (s)f (x) = f (y)p(s, x, dy) = Ex [f (Xs )].
S
Consequently, Z
T (s)f (Xt ) = f (y)p(s, Xt , dy) = E[f (Xt+s )|FtX ].
S
Exercise 4.9. The family of operators {T (t); t ≥ 0} satisfies
1. T (0)f = f
2. T (s + t)f = T (s)T (t)f
for all bounded measurable f .
4 MARKOV PROCESSES 35
lim T (t)f = f.
t→0+
Exercise 4.11. Let A be a bounded linear operator and define, for t ≥ 0, the linear operator
∞
X 1 k k
exp tA = t A , exp 0 = I.
k!
k=0
Then, exp tA is a strongly continuous semigroup and || exp tA|| ≤ exp t||A||.
Proposition 4.12. Let T be a strongly continuous contraction semigroup. Then, for each f ∈ L, the
mapping
t → T (t)f
is a continuous function from [0, ∞) into L.
Proof. For t ≥ 0 and h > 0,
||T (t + h)f − T (t)f || = ||T (t)(T (h)f − f )|| ≤ ||T (t)|| ||T (h)f − f || ≤ ||T (h)f − f ||.
For 0 ≤ h ≤ t,
||T (t − h)f − T (t)f || = ||T (t − h)(T (h)f − f )|| ≤ ||T (t − h)|| ||T (h)f − f || ≤ ||T (h)f − f ||.
Now, let h → 0.
2. L × L is a Banach space under the norm ||(f, g)|| = ||f || + ||g||. A is a closed operator if its graph is
a closed subspace of L × L.
Note that if D(A) = L, then A is closed.
4 MARKOV PROCESSES 36
The domain of G is the subspace of all f ∈ L for which the limit above exists.
Exercise 4.15. 1. Let L be the bounded real-valued functions on N. For the Poisson process with param-
eter λ, show that the generator is
Exercise 4.16. 1. If u : [a, b] → L is continuous and ||u|| is Riemann integrable, then u is integrable and
Z b Z b
|| u(s) ds|| ≤ ||u(s)|| ds.
a a
2. Let B be a closed linear operator on L and assume that u : [a, b] → L is continuous, that u(s) ∈ D(B)
for all s ∈ [a, b], and that both u and Bu are Riemann integrable. Then
Z b Z b Z b
u(s) ds ∈ D(B) and B u(s) ds = Bu(s) ds.
a a a
d
T (t)f = GT (t)f = T (t)Gf. (4.3)
dt
Proof. 1. Observe that (T (h) − I)/h is closed for all h > 0. Thus,
Z t Z t Z t+h Z h
1 1 1 1
(T (h) − I) T (s)f ds = (T (s + h)f − T (s)f ) ds = T (s)f ds − T (s)f ds.
h 0 h 0 h t h 0
d+
T (t)f = GT (t)f = T (t)Gf.
dt
To check the left derivative, use the identity
1
(T (t − h)f − T (t)f ) − T (t)Gf = T (t − h)(Gh − G)f + (T (t − h) − T (t))Gf,
−h
valid for h ∈ (0, t].
3. This is a consequence of part 2 in this proposition and part 3 in the lemma above.
T (s)f (x) = f (x + s)
D(G) is dense in L.
4 MARKOV PROCESSES 38
fn → f and Gfn → g as n → ∞.
Then, by (4.3),
Z t
T (t)fn − fn = T (s)Gfn ds for each t > 0.
0
Thus, by letting n → ∞, we obtain
Z t
T (t)f − f = T (s)g ds for each t > 0.
0
f ∈ D(G) and Gf = g.
Definition 4.20. For a strongly continuous contraction semigroup T , define the resolvent,
Z ∞
R(λ)g = e−λs T (s)g ds.
0
Exercise 4.21. If X is a time homogenuous Markov process with semigroup T and τλ is an independent
exponential random variable with parameter λ, then
λR(λ)g = Eg(Xτλ ).
Proposition 4.22. Let T be a strongly continuous contraction semigroup with generator G, then for all
λ > 0, (λI − G)−1 is a bounded linear operator on L and
for all g ∈ L.
Proof. Fix λ > 0 and choose g ∈ L. Then
Z ∞
1
||R(λ)g|| = e−λs ||T (s)g|| ds ≤ ||g||.
0 λ
(λI − G)R(λ)g = g.
or
R(λ)(λI − G)g = g.
Thus, (λI − G) is one-to-one and its left inverse is R(λ)
Definition 4.23. ρ(G) = {λ; (λI − G) has a bounded inverse} is called the resolvent set.
Thus, if G is the generator of a strongly continuous contraction semigroup, then its resolvent set ρ(G)
contains (0, ∞).
Remark 4.24. Let τλ be an exponential random variable with parameter λ, then as λ → ∞, τλ converges
in distribution to the degenerate random varibale that takes on the value 0 with probability 1. This motivates
the following.
Proposition 4.25. Let {R(λ); λ ≥ 0} be the resolvent for a strongly continuous contraction semigroup with
generator G, then
lim λR(λ)f = f.
λ→∞
Thus, the formula holds for the dense set f ∈ D(G). Now, use the fact that ||λR(λ) − I|| ≤ 2 for all λ > 0
and approximate.
Exercise 4.26. 1. (resolvent identity) Let λ, µ > 0, then
R(µ) − R(λ)
R(λ)R(µ) = R(µ)R(λ) = .
λ−µ
Remark 4.28. If A is dissipative, then divide (4.5) by λ to obtain that ||f − Af /λ|| ≥ ||f || or
and {fn ; n ≥ 0} is a Cauchy sequence. Thus, there exists f ∈ L such that fn → f and hence
Afn → λf − h.
(λI − A)fn → λf − g.
Because A is dissipative,
||(λI − A)(fn − f0 )|| ≥ λ||fn − f0 ||,
and fn → f0 . Hence, f = f0 ∈ D(A) and Af = g. This shows that A is closed.
Lemma 4.31. Let A be a dissipative closed linear operator on L. Set ρ+ (A) = ρ(A) ∩ (0, ∞). If ρ+ (A) 6= ∅,
then ρ+ (A) = (0, ∞)
4 MARKOV PROCESSES 41
Proof. The condition is equivalent to the statement that ρ+ (A) is both open and closed in (0, ∞). We have
shown that ρ(A) is open.
Suppose that {λn ; n ≥ 1} ⊂ ρ+ (A) with λn → λ > 0. Given g ∈ L, define, for each n,
Because A is dissipative
|λ − λn |
||gn − g|| = ||((λI − A) − (λn I − A))(λn I − A)−1 g|| = ||(λ − λn )(λn I − A)−1 g|| ≤ ||g||.
λn
Thus, gn → g. Consequently, R(λI − A) is dense in L. We have shown that this range is closed, thus
R(λI − A) = L. Because λI − A is one-to-one and ||(λI − A)−1 || ≤ λ−1 , we have that λ ∈ ρ+ (A). Thus,
ρ+ (A) is closed.
Definition 4.32. Let G be a dissipative closed linear operator on L, and suppose that D(G) is dense in L
and that (0, ∞) ⊂ ρ(G). Then the Yosida approximation of G is
Proof. Because
I = (λI − G)R(λ) on L, Aλ = λ2 R(λ) − λI on L.
Because
R(λ)(λI − A) = I on D(G), Aλ = λR(λ)G on D(G).
Note that 2
||Tλ (t)|| ≤ e−tλ || exp(tλ2 R(λ))|| ≤ e−tλ etλ ||R(λ)||
≤ 1.
proving part 1.
Part 2 follows from the first identity above and the resolvent identity.
Part 3 follows from the second identity above and the fact that for f ∈ D(G), λR(λ)f → f as λ → ∞ as
λ → ∞.
4 MARKOV PROCESSES 42
Lemma 4.34. For a pair of commuting bounded operators B and C on L assume that || exp(tB)|| ≤ 1 and
|| exp(tC)|| ≤ 1 for all t ≥ 0. Then
Proof.
Z t
d
exp(tB)f − exp(tC)f = (exp(sB) exp((t − s)C))f ds
0 dt
Z t
= exp(sB)(B − C) exp((t − s)C)f ds
0
Z t
= exp(sB) exp((t − s)C)(B − C)f ds.
0
lim Tλ (t)f
λ→∞
exists for all t ≥ 0 uniformly on bounded intervals. Call this limit T (t)f . Because D(G) is dense, this limit
holds for all f ∈ L.
Claim 1. T is a strongly continuous contraction semigroup.
The fact that T (0)f = f and that T (t) is a contraction is immediate. Use the identity
T (s + t)f − T (s)T (t)f = (T (s + t) − Tλ (s + t))f + Tλ (s)(Tλ (t) − T (t))f + (Tλ (s) − T (s))T (t)f
4 MARKOV PROCESSES 43
Xn Z tj
≤ ||u()|| + ||Au(tj ) − Au(s)|| ds.
j=1 tj−1
4 MARKOV PROCESSES 44
The first sum is negative by the dissipativity of A. Now use the continuity of Au and u and let max(tj −
tj−1 ) → 0. Then, let → 0.
Proposition 4.37. Let T and S be strongly continuous contraction semigroups on L with generators GT
and GS . If these two generators are equal, the T = S.
Proof. Use the lemma above with
u(t) = (T (t) − S(t))f
and note that u(0) = 0 and Au(s) = 0 for all s.
Definition 4.38. Call a linear operator A on L closable if it has a closed linear extension. The closure Ā
of A is the minimal closed linear extension of A.
Often the generator G is an unbounded operator. In these instances, establishing D(G) can be a tedious
process. Thus, we look for an alternative form for the Hille-Yosida theorem.
Lemma 4.39. Let G be a dissipative linear operator on L with D(G) dense in L. Then G is closable and
R(λI − G) = R(λI − Ḡ) for every λ > 0.
Proof. For the first assertion, we show that
||(λI − G)gm − λg|| = lim ||(λI − G)(gm + λfn )|| ≥ lim λ||gm + λfn || = λ||gm ||
n→∞ n→∞
for every λ > 0. Now divide by λ and let λ → ∞ to obtain, for each m, ||gm − g|| ≥ ||gm ||. Letting m → ∞
yields g = 0.
The inclusion R(λI − Ḡ) ⊂ R(λI − G) is straightforward. The reverse inclusion follows from the fact
that dissipative linear operator G̃ is closed if and only if R(λI − G̃) is closed.
Theorem 4.40. A linear operator G on L is closable and its closure Ḡ is the generator of a strongly
continuous contraction semigroup on L if and only if
1. D(G) is dense in L.
2. G is dissipative.
3. R(λ0 I − G) is dense in L for some λ0 > 0
Proof. By the lemma above G satisfies 1-3 if and only if G is closable and Ḡ satisfies 1-3 in the Hille-Yosida
theorem.
Definition 4.41. Let A be a closed linear operator on L. A subspace D of D(A) is called a core for A if A
is the closure of the restriction of A to D.
Proposition 4.42. Let G be the generator of a strongly continuous contraction semigroup on L. Then a
subspace D of D(G) is a core for G if and only if D is dense in L and R(λ0 I − G) is dense in L for some
λ0 > 0.
4 MARKOV PROCESSES 45
Definition 4.43. Let S be locally compact and let C0 (S) be the Banach space of continuous functions that
vanish at infinity with norm
||f || = sup{|f (x)| : x ∈ S}.
Definition 4.44. Call an operator on a function space positive if it maps nonnegative functions to nonneg-
ative functions.
To establish positivity, we will use the following proposition.
Proposition 4.45. Let T be a strongly continuous contraction semigroup on L with generator G and resol-
vent R. For M ⊂ L, let
ΛM = {λ > 0; λR(λ) : M → M }.
If M is closed and convex and ΛM is unbounded, then for each t ≥ 0,
T (t) : M → M.
Because M is closed and convex, λ ∈ ΛM implies (0, λ] ⊂ ΛM . Thus, ΛM = (0, ∞). Use the Yosida
approximation
∞
X (tλ)n
Tλ (t) = e−tλ exp(tλ(λR(λ))) = e−tλ (λR(λ))n
n=0
n!
to conclude that
Tλ (t) : M → M.
Now use the fact that M is closed and that Tλ (t)f converges to to T (t)f as λ → ∞.
Definition 4.46. 1. Call a semigroup conservative if there exist a sequence fn ∈ C0 (S) that is bounded
in norm and converges pointwise to 1, and
lim T (t)fn = 1.
n→∞
for all t ≥ 0
2. Call a conservative strongly continuous positive contraction semigroup T on C0 (S) a Feller semigroup.
Exercise 4.47. 1. If T is conservative, then A1 = 0
2. If A1 = 0, then T is conservative.
4 MARKOV PROCESSES 46
Theorem 4.51. Let S be locally compact. A linear operator G on C0 (S) is closable and its closure Ḡ is the
generator of a positive strongly continuous contraction semigroup on C0 (S) if and only if
1. D(G) is dense in C0 (S).
2. G satisfies the positive maximum principle.
3. R(λ0 I − G) is dense in C0 (S) for some λ0 > 0.
Proof. The necessity of conditions 1 and 3 follows from the theorem above. To check the neccessity of 2, fix
f ∈ D(G) and x̃ ∈ S so that sup{f (x) : x ∈ S} = f (x̃) ≥ 0. Then, for each t ≥ 0,
and Gf (x̃) ≤ 0.
Conversely, suppose G satisfies conditions 1-3. Because condition 2 implies that G is dissipative, Ḡ
generates a strongly continuous contraction semigroup T . To prove that T is positive, we note, by the
proposition above, it suffices to prove that R(λ) maps nonnegative functions to nonegative functions.
Because R(λ0 I − G) is dense for some λ0 > 0, and G is dissipative, R(λI − Ḡ) = C0 (S) all λ > 0. Thus,
it is equivalent to show that for f ∈ D(G), and λ > 0,
To establish the contrapositive of this statement, choose f ∈ D(Ḡ) so that inf{f (x) : x ∈ S} < 0. Thus,
there exist {fn ; n ≥ 0} ∈ D(G) so that
Because Ḡ is dissipative, we have that fn → f as n → ∞. Let x̃n and x̃ be the points in which fn and f ,
respectively take on their minimum values. Then
inf{(λI − Ḡ)f ; x ∈ S} ≤ lim inf (λI − Ḡ)fn (x̃n ) ≤ lim inf λfn (x̃n ) = λf (x̃) < 0.
n→∞ n→∞
Exercise 4.52. Let X1 and X2 be independent Markov processes on a common state space S. Is (X1 , X2 )
a Markov process? If so, determine its generator?
X is called strong Markov with respect to {Ft ; t ≥ 0} if X is strong Markov for all almost finite Ft -
stopping times.
In reviewing the results on Lévy processes and its regularity properties, we obtain the following.
Proposition 4.55. Let X be a Ft -progressive, Ft Markov process with associated transition function P .
The X is strong Markov at all discrete almost surely finite Ft -stopping times.
Now, using the essentially the same proof we had for Lévy processes, we have
Proposition 4.56. Let X be a right continuous Ft -progressive, Ft Markov process with associated transition
function P . Then X is strong Markov at all almost surely finite Ft+ -stopping times.
Theorem 4.57. Let X be a Markov process with time homogenous transition function P . Let Y be a bounded
measurable process and assume that and let τ be a FtX -stopping time so that X is strong Markov at τ + t
for all t ≥ 0. Then
Eα [Yτ ◦ θτ |Fτ ] = φτ (Xτ ) on {τ < ∞}
where
φs (x) = Ex Ys .
4 MARKOV PROCESSES 48
Proof. By the standard machine, we need only consider functions of the form
n
Y
Ys = f0 (s) fk (Xtk ), 0 < t1 < · · · < tn .
k=1
Then, Z
φs (x) = f0 (s) φ(x1 )f1 (x1 )p(t1 , x, dx1 )
On {τ < ∞},
n
Y
Eα [Yτ ◦ θτ |Fτ ] = Eα [f0 (τ ) fk (Xτ +tk )|Fτ ]
k=1
n
Y
= f0 (τ )Eα [Eα [ fk (Xτ +tk )|Fτ +t1 ]f1 (Xτ +t1 )|Fτ ]
k=2
Using essentially the same proof as in the case of Lévy processes, we have the following.
Theorem 4.58 (Blumenthal 0-1 law). Let X be a right continuous Ft -progressive, Ft Markov process. Then
X
for each x ∈ S every event in F0+ has probability 0 or 1.
X
Corollary 4.59. Let τ be an Ft+ -stopping time and fix x ∈ S. Then either
Px {τ = 0} = 1 or Px {τ > 0} = 1.
4 MARKOV PROCESSES 49
is an FtX -martingale.
Proof. For each t, u ≥ 0
Z t Z t+u
f
E[Mt+u |FtX ] = T (u)f (Xt ) − Gf (Xs ) ds − T (s − t)Gf (Xt ) ds
0 t
Z u Z t
= T (u)f (Xt ) − T (s)Gf (Xt ) ds − Gf (Xs ) ds
0 0
Z t
= f (Xt ) − Gf (Xs ) ds = Msf
0
Exercise 4.61. Let X be a right continuous Markov process with generator G and assume that
Z t
f (Xt ) − g(Xs ) ds
0
Theorem 4.62 (Dynkin formula). In addition to the conditions above, let τ be a stopping time with Eα τ <
∞, then Z τ
Eα f (Xτ ) = Eα f (X0 ) + Eα [ Gf (Xs ) ds]
0
Analogous to the situation for discrete time Markov chains, we make the following definitions.
Definition 4.63. Let G be the generator for a time homogeneous Markov process X and let f ∈ D(G).
Then call
1. f harmonic if Gf = 0.
2. f superharmonic if Gf ≤ 0.
3. f subharmonic if Gf ≥ 0.
Remark 4.64. 1. If f harmonic, then f (Xt ) is a martingale.
2. If f superharmonic, then f (Xt ) is a supermartingale.
4 MARKOV PROCESSES 50
Assume that the first entrance times τDj , j = 0, 1 are stopping times and τ is their minimum. Then if
τ has finite mean,
h(x) = Ex h(Xτ ) = P {τD0 > τD1 }.
Mtf = f (Xt ) + t
f (x) = Ex τD .
Lemma 4.66. Assume that g ≥ 0 and let R(λ) be the resolvent for a positive continuous contraction
semigroup T associated to a Markov process X. Then
e−λt R(λ)g(Xt )
X
is a non-negative Ft+ -supermartingale.
Proof. For t > s, use the fact that the semigroup and the resolvent commute to obtain that
Let’s continue this line analysis. Consider the bounded random variable,
Z ∞
Y = e−λs g(Xs ) ds
0
Then Z t
Y = e−λs g(Xs ) ds + e−λt Y ◦ θt
0
and
Ex Y = R(λ)g(x).
4 MARKOV PROCESSES 51
Now, because Doob’s martingale is uniformly integrable, any stopping time τ satisfies the sampling
integrability conditions for Z λ,g and therefore,
Z τ
R(λ)g(x) = Ex [ e−λs g(Xs ) ds] + Ex [e−λτ R(λ)g(Xτ )].
0
Now, let f = (λI − G)g, then the Doob’s martingale above becomes
Z t
Ctλ,f = e−λt f (Xt ) + e−λs (λI − G)f (Xs ) ds
0
fλ (x) = Ex e−λτD .
Theorem 4.67. Let X be a measurable Ft -adapted process and let f and g be bounded, measurable functions,
inf x f (x) > 0. Assume that
Z t
Yt = f (Xt ) − g(Xs ) ds
0
is an Ft martingale. Then Z t
g(Xv )
f (Xt ) exp − dv
0 f (Xv )
is a martingale.
Proof. Let Z t
g(Xv )
Vt = exp − dv .
0 f (Xv )
4 MARKOV PROCESSES 52
For the double integral, reverse the order of integration, then we have, upon integrating the s variable that
Z t Z s
g(Xv ) t
− g(Xu ) exp − dv du
0 0 f (Xv ) u
Z t Z t Z t Z u
g(Xv ) g(Xv )
=− g(Xu ) exp − dv du + g(Xu ) exp − dv du
0 0 f (Xv ) 0 0 f (Xv )
is an Ft martingale. Then Z t
f (Xt ) − g(Xs ) ds
0
is an Ft martingale.
2. Assume that Z t
f (Xt ) − g(Xs ) ds
0
is an Ft martingale. Then
Z t
e−λt f (Xt ) + e−λs (λf (Xs ) − g(Xs )) ds
0
is an Ft martingale.
4 MARKOV PROCESSES 53
Ex f (Xh ) − f (x)
Gf (x) = lim , f ∈ D(G).
h→0 h
To relate these two concepts, write
X X
Ex f (Xh ) = f (y)g(x, y)h + f (x)(1 − g(x, y)h) + o(h)
y6=x y6=x
X
Ex f (Xh ) − f (x) = g(x, y)(f (y) − f (x))h + o(h)
y6=x
X
Gf (x) = g(x, y)(f (y) − f (x))
y∈S
Thus, G can be written as an infinitesimal transition matrix. The xy-entry x 6= y is g(x, y). The diagonal
entry X
g(x, x) = − g(x, y).
y6=x
Thus, the off-diagonal entries of the matrix G are non-negative and the row sum is 1.
τ1 = inf{t ≥ 0; Xt 6= X0 }
Theorem 4.71 (structure theorem for pure jump Markov processes). Under Px , τ1 and Xτ1 are independent
and there is a B(S) measurable function λ on S such that
Proof. Set
ex (t+s) = Px {τ1 > t+s} = Px {τ1 > t+s, Xt = x, τ1 > t} = Px {τ1 > t+s|Xt = x, τ1 > t}Px {Xt = x, τ1 > t}
Note that
P {τ1 > t + s|Xt = x, τ1 > t} = Px {τ1 > t + s|Xt = x} = Px {τ1 > s} = ex (s).
and that Px {Xt = x, τ1 > t} = Px {τ1 > t} = ex (t). Thus,
ex (t + s) = ex (s)ex (t)
or for some λ ∈ [0, ∞], ex (t) = exp(−λ(x)t). The function λ is measurable because Px {τ1 > t} is measurable.
Let τ1 (t) be the first exit from state x after time t. Then, for B ∈ B(S), x ∈
/ B,
Px {Xτ1 ∈ B, τ1 > t} = Px {Xτ1 (t) ∈ B, τ1 > t, Xt = x} = Px {Xτ1 (t) ∈ B|τ1 > t, Xt = x}Px {τ1 > t, Xt = x}.
Ex f (Xh ) = Ex [f (Xh )|τ1 > h]Px {τ1 > h} + Ex [f (Xh )|τ1 ≤ h]Px {τ1 ≤ h}
Z
−λ(x)h
= f (x)e + f (y)µ(x, dy)(1 − e−λ(x)h ) + o(h)
Thus, Z
Ex f (Xh ) − f (x) = (1 − e−λ(x)h ) f (y)µ(x, dy) − f (x) + o(h)
and Z
Gf (x) = λ(x) (f (y) − f (x))µ(x, dy).
In the case of a countable set of states, set µ(x, {y}) = T (x, y).
X
Gf (x) = λ(x) T (x, y)(f (y) − f (x))
y∈S
g(x, x) = −λ(x)
4 MARKOV PROCESSES 55
and for y 6= x
g(x, y)
g(x, y) = λ(x)T (x, y) or T (x, y) =
λ(x)
and
Xt = Yk , τk ≤ t < τk+1 .
Note that we can allow λ(x) = 0. In this case, once X arrives to x it remains there. Thus, if the Markov
process arrives at the site x is remains there an exponential length of time, parameter λ(x) and then jumps
to a new site according to the transition function µ.
Let us consider the alternative representation for G. Define
λ(x) λ(x)
µ̃(x, B) = 1 − δx (B) + µ(x, B).
λmax λmax
R
Exercise 4.75. If G̃f (x) = λmax S
(f (y) − f (x)) µ̃(x, dy), then G = G̃
Write Z
T f (x) = f (y) µ̃(x, dy)
S
G = λmax (T − I).
Let {Ỹk ; k ≥ 0} be a Markov chain on S with initial distribution α and transition function µ̃ and let N
be an independent Poisson process with parameter λmax , and define
X̃t = ỸNt .
E[f (Ỹk+Nt ); A ∩ B ∩ {Nt = m}] = E[f (Ỹk+m ); A ∩ B ∩ {Nt = m}] = P (A ∩ {Nt = m})E[f (Ỹk+m ); B]
= P (A ∩ {Nt = m})E[T k f (Ỹm ); B] = E[T k f (X̃t ); A ∩ B ∩ {Nt = m}]
However, sets of the form A ∩ B ∩ {Nt = m} are closed under finite intersection, and the claim follows from
the Sierpinski class theorem.
4 MARKOV PROCESSES 57
Now, use the independence of the increments for a Poisson process, to obtain that
E[f (X̃t+s )|Ft ] = E[f (ỸNt+s )|Ft ] = E[f (ỸNt+s −Nt +Nt )|Ft ]
∞ ∞
X (λmax s)k X (λmax s)k k
= e−λmax s E[f (Ỹk+Nt )|Ft ] = e−λmax s T f (X̃t ) = T (s)f (X̃t )
k! k!
k=0 k=0
Because G = G̃, X is a Markov process whose finite dimensional distributions agree with X̃.
Thus, all of the examples of generators with bounded rates are generators for a semigroup T for a Markov
process on C0 (S).
Exercise 4.77. For a two state Markov chain on S = {0, 1} with generator
−λ λ
A= ,
µ −µ
show that
µ λ −(λ+µ)t
P (t, 0, {0}) = + e .
λ+µ λ+µ
λ(x) is some constant λ, the rate for the Poisson process. If ν is the distribution of the Yn , then
Gf (x, y) = rx(N − x − y)(f (x + 1, y) − f (x, y)) + µx(f (x − 1, y + 1) − f (x, y)) + γy(f (x, y − 1) − f (x, y)).
12. For the Moran model with mutation S = {0, 1, · · · , N }, consider a genetic model with two alleles, A1
and A2 Xt - number of alleles of type A1 at time t. The process remains at a site for an exponential
length of time, parameter λ, an individual is chosen at random to be replaced. A second individual
is chosen at random to duplicate. A mutation may also occur at birth. The mutation A1 → A2 has
probably κ1 and the mutation A2 → A1 has probably κ2 The generator
x x x
Gf (x) = λ(1 − ) (1 − κ1 ) + (1 − )κ2 (f (x + 1) − f (x))
N N N
x x x
+λ (1 − )(1 − κ2 ) + κ1 (f (x − 1) − f (x))
N N N
13. For a random walk on the lattice S = Zd , Xt - position of the particle at time t. The particle at a
site x remains at x for an exponential length of time, parameter λ(x), and then moves according to the
kernel p(x, y).
Thus, X
p(x, y) = 1
y
Exercise 4.79. For the compound Poisson process, if {Yn ; n ≥ 1} is a Bernoulli sequence, the X is a
Poisson process with parameter λp
To find harmonic functions h̃ for birth and death processes, note that
Assume that h̃(0) = 0 and h̃(1) = 1 and all of the λx > 0. Then any other non-constant harmonic
function is a linear function of h̃. Then
y−1 y−1 x
X XY µx̃
h̃(y) = (h̃(x + 1) − h̃(x)) = .
x=0 x=1 x̃=1
λx̃
Thus, for a simple birth and death process and for a linear birth and death process,
y−1
X µ x 1 − (µ/λ)y
h̃(y) = = .
x=0
λ 1 − (µ/λ)
Exercise 4.80. For a birth and death process, let τ = min{τ0 , τN }, for x ∈ {0, . . . , N }.
1. Find Ex τ .
2. Find Ex e−λτ , λ > 0.
3. Consider what happens as N → ∞.
Exercise 4.81. 1. For a simple birth and death process, X, find Ex Xt and Varx (Xt ).
2. For a linear birth and death process, X, find Ex Xt and Varx (Xt )
η:Λ→F
where F is a finite set. Examples of lattices are Zd , Z/M Z)d , hexagonal lattice, and regular trees. Examples
of F are
{0,1} = {vacant, occupied}
{-1,1}= {spin up, spin down}
{0,1,· · · ,k} = {vacant, species 1, species 2, · · · , species k}
{∆, 0, · · · , N} = {vacant, 0 individuals having allele A, · · · , N individual having allele A}
4 MARKOV PROCESSES 60
Example 4.82 (interacting particle systems). 1. Exclusion Process (Spitzer, 1970). F = {0, 1}, Λ =
Zd .
If initially we have one particle at x0 , i.e.,
(
1, if x = x0 ,
η0 (x) =
0, if x 6= x0 .
Set Xt = ηt−1 ({1}). Then {Xt ; t ≥ 0} is the random walk on the lattice.
When there are many particles, then each executes an independent random walk on the lattice excluding
transitions that take a particle to an occupied site. The generator
X
GF (η) = λ(η, x)p(x, y)η(x)(1 − η(y))(F (η xy ) − F (η))
x,y
where
η(x),
if z = y,
η xy (z) = η(y), if z = x,
η(z), otherwise.
where (
1 − η(x), if z = x,
ηx (z) =
η(z), otherwise.
3. Voter Model (Clifford and Sudbury, 1973, Holley and Liggett, 1975). F = {0, 1}={no, yes}, Λ = Zd .
An individual changes opinion at a rate proportional to the number of neighbors that disagrees with the
individual’s present opinion. The generator
X
GF (η) = λ(η, x)(F (ηx ) − F (η))
x,y
where
λ X
λ(η, x) = I{η(y)6=η(x)} .
2d
{y;|y−x|=1}
4 MARKOV PROCESSES 61
where ( P
λ {y;|y−x|=1} η(y), if η(x) = 0,
λ(η, x) =
1, if η(x) = 1.
Thus, λ is the relative infection rate.
5. For the stochastic Ising model, F = {−1, 1}. Λ = Zd .
X X
GF (η) = exp −β η(x)η(y) (F (η̄x ) − F (η))
x,y y:|x−y|=1
where (
−η(x), if z = x,
η̄x (z) =
η(z), otherwise.
σn = inf{t > 0; Xt ∈
/ Kn }.
ζ = lim σn .
n→∞
If this limit is finite, then we say that an explosion occurs. In these circumstances, we have that the transition
function fails to satisfy P (t, x, S) = 1 for all t > 0, and the semigroup T is no longer conservative.
4 MARKOV PROCESSES 62
Xt = ∆ for t ≥ ζ.
Theorem 4.83. Let S be locally compact and separable and let T be a strongly continuous positive contraction
semigroup on C0 (S). For each t ≥ 0, define the operator T ∆ (t) on C(S ∆ ) by
Claim 1. T ∆ is positive. In other words, given c ∈ R+ and f ∈ C0 (S), c+f ≥ 0 implies that c+T (t)f ≥ 0.
Because T is positive,
T (t)(f + ) ≥ 0 and T (t)(f − ) ≥ 0.
Hence,
−T (t)f ≤ T (t)(f − ) and so (T (t)f )− ≤ T (t)(f − ).
Because T (t) is a contraction, we obtain
Therefore,
T (t)(f − )(x) ≤ c and so c + T (t)f ≥ 0.
Claim 2. T ∆ is a contraction.
Because T ∆ is positive,
|T ∆ (t)f (x)| ≤ T ∆ (t)||f || = ||f ||
and ||T ∆ (t)|| ≤ 1.
With this modification, the condition of being a Feller semigroup states that the distribution of the
Markov process at time t depends continuously on the initial state. In other words, for f ∈ C(S ∆ ),
τx = σ 1 + · · · + σ x
where σx̃ is the length of time between the arrival of X to x̃ − 1 and x̃.
Then, the random variables {σx̃ ; x̃ ≥ 1} are independent exponential random variables with parameters
{λx̃−1 ; x̃ ≥ 1}. Thus, the Laplace transform
λx̃−1
E[e−uσx̃ ] = ,
u + λx̃−1
and !−1
x x
Y λx̃−1 Y u
E[e−uτx ] = = (1 + ) .
λx̃−1 + u λx̃−1
x̃=1 x̃=1
X̃t ≤ Xt
Because ||λR(λ)|| ≤ 1, we have that µλ (x, S) ≤ 1. However, if G is conservative, we can choose f (x) = 1/λ
to obtain
1 = µλ (x, S).
where τλ is an exponential random variable, parameter λ, and independent from X. Thus, µλ (x, ·) is the
distribution of Xτλ under Px .
Exercise 4.84. let D be a dense subset of C0 (S). Then a function x ∈ DS [0, ∞) if and only if f (x) ∈
DR [0, ∞) for every f ∈ D.
X
Theorem 4.85. Let X be a Ft+ -Feller process, then X has a DS [0, ∞)-valued modification.
Proof. Let X have generator G and let f ∈ D(G), then the martingale
Z t
f
Mt = f (Xt ) − Gf (Xs ) ds
0
Theorem 4.86. Let S be locally compact and separable. Let T be a positive strongly continuous contraction
semigroup on C0 (S) and define T ∆ as above. Let X be a Markov process corresponding to T ∆ with sample
paths in DS ∆ [0, ∞) and let
ζ = inf{t ≥ 0 : Xt = ∆ or Xt− = ∆}.
Then on the set {ζ < ∞},
Xζ+s = ∆ for all s ≥ 0.
Proof. By the Urysohn lemma, there exists f ∈ C(S ∆ ) such that f > 0 on S and f (∆) = 0. Then the
resolvent R1∆ f > 0 on S and R1∆ f (∆) = 0. Note that e−t R1∆ f (Xt ) is a supermartingale. Because a right
continuous nonnegative supermartingale is zero after its first contact with 0, the theorem follows.
4 MARKOV PROCESSES 65
Proposition 4.87. If, in addition to the conditions above, if the extention the semigroup T to the bounded
measurable functions is conservative, then α(S) = P {X0 ∈ S} = 1 implies that Pα {X ∈ DS [0, ∞) = 1}.
Proof. Let G∆ be the generator of T ∆ , then G∆ IS = 0, thus
Z t
Eα IS (Xt ) = Eα IS (X0 ) + Eα [ G∆ IS (Xs ) ds] = Eα IS (X0 ) + 0.
0
or
Pα {ζ > t} = Pα {Xt ∈ S} = Pα {X0 ∈ S} = 1.
Theorem 4.88. Let S be locally compact and separable. Let T be a Feller semigroup on C0 (S). Then for
each probability measure α on S, there exists a Markov process X with initial distribution α and sample paths
X
in DS [0, ∞). Moreover, X is strong Markov with respect to the filtration {Ft+ ; t ≥ 0}.
Proof. Let G be the generator for T and let Gλ , λ > 0, be its Yosida approximation. Then Tλ is the
semigroup for a jump Markov process X λ . Moreover, for every t ≥ 0 and f ∈ C0 (S),
Let Tλ∆ and T ∆ be the corresponding semigroups on C(S ∆ ). Then, for each f ∈ C(S ∆ ),
Now f → T ∆ (t)f (x) α(dx) is a positive linear operator that maps the constant function 1 to the value
R
1, and thus, by the Riesz representation theorem is integration against a probability measure νt . This gives
the one dimensional distributions of X. The proof that the finite dimensional distributions converge follows
from a straightforward induction argument.
X
To check the strong Markov property, let τ be a discrete Ft+ -stopping time, then for each > 0, and
X
A ∈ Fτ , A ∩ {τ = t} ∈ Ft+ . Thus, for s > 0,
Now use the right continuity of X and the strong continuity of T to conclude that
We can obtain the result for an arbitrary stopping time by realizing it as the decreasing sequence of
discrete stopping times, use the identity above, the right continuity of X, and the strong continuity of T .
4 MARKOV PROCESSES 66
Exercise 4.89. Complete the proof above by supplying the induction argument to show that the finite di-
mensional distributions converge.
In summary, we have
Theorem 4.90. Let S be locally compact and separable. Let G be a linear operator on C0 (S) satisfying 1-3
of the Hille-Yosida theorem and let T be a strongly continuous postive contraction semigroup generated by
Ḡ. Then for each x ∈ S, there exists a Markov process X corresponding to T with initial distribution δx and
with sample paths in DS [0, ∞) if and only if G is conservative.
Rt
Set σ0 = inf{t > 0 : 0
1/c(Xτs ) ds = ∞}, then for t < σ0 ,
Z t Z τt
c(Xs ) 1
t= ds = du
0 c(Xs ) 0 c(Yu )
τt0 = c(Yτt ), τ0 = 0.
Because c(Y ) is bounded on bounded time intervals, the solution to the ordinary differential equation
has a unique solution. Now,
Xt = Yτt .
Theorem 4.93. Let Y be a DS [0, ∞)-valued Markov process with generator G on C0 (S). If a unique solution
to
Xt = YR t c(Xs ) ds
0
is an FτYt -martingale.
Thus, f in in the domain of the generator of the Markov process X and this generator is cA.
Exercise 4.94. Verify that Y in the theorem above is a Marlov process.
4.8.2 Killing
We can choose to have a process X move to the state ∆. The rate that this happens can be spatially
dependent. Thus, we define the hazard function
1
k(x) = lim Px {Xh = ∆}.
h→0 h
The analysis follows the previous analysis for the case k constant. Write the additive functional
Z t
Kt = k(Xs ) ds
0
and define Z ∞
Y = e−Ks f (Xs ) ds, Rk f (x) = Ex Y.
0
4 MARKOV PROCESSES 68
Then Z t
Y = e−Ks f (Xs ) ds + e−Kt Y ◦ θt
0
Again, we have Doob’s martingale
Z t
Ztk,f = E[Y |FtX ] = e−Ks f (Xs ) ds + e−Kt E[Y ◦ θt |FtX ]
0
Z t
= e−Ks f (Xs ) ds + e−Kt Rk f (Xt )
0
Rk (k − G)f = f.
Ex f (X̃t ) = Ex [f (Xt )I{Kt <ξ} ] = Ex [f (Xt )Ex [I{Kt <ξ} |FtX ]] = Ex [e−Kt f (Xt )].
Then T is a positive continuous constraction semigroup with generator G − k. Thus, X̃ is Markov process
with state space S ∆ and generator G − k.
4 MARKOV PROCESSES 69
Corollary 4.97. (Feynman-Kac formula) Let X be a Feller process on S and generator G. Then the
equation
∂u
= Gu − ku, u(0, x) = f (x)
∂t
has solution
u(t, x) = Ex [e−Kt f (Xt )].
E[ZL|G]
EQ [Z|G] = .
E[L|G]
Lemma 4.99. Assume Q|Ft << P |Ft , and assume that Lt is the Radon-Nikodym derivative, then Z is a
Q-martingale if and only if ZL is a P -martingale.
Proof. Because L is an adapted process, Z is adapted if and only if ZL is adapted. To show that L is a
martingale, choose A ∈ Ft , then Q(A) = E[Lt ; A]. However, A ∈ Ft+s and therefore
By Bayes formula,
E[Zt+s Lt+s |Ft ] E[Zt+s Lt+s |Ft ]
EQ [Zt+s |Ft ] = = .
E[Lt+s |Ft Lt
Thus,
E[Zt+s Lt+s |Ft ] − Zt Lt
EQ [Zt+s |Ft ] − Zt = .
Lt
and, consequently,
EQ [Zt+s |Ft ] = Zt if and only if E[Zt+s Lt+s |Ft ] − Zt Lt .
4 MARKOV PROCESSES 70
Note that L is a positive mean one martingale. In addition, is L−1 is Q-integrable, the L−1 is a Q-
martingale and P |Ft << Q|Ft
For X, a time homogeneous Markov process with generator G, we write g = Gf for f ∈ D(G), with
inf x f (x) > 0, recall that Z
f (Xt ) g(Xv )
exp − dv
f (X0 ) f (Xv )
is a mean 1 local martingale.. Write f (x) = exp w(x) and Gw = e−w Gew , then this local martingale becomes
Z t
Lt = exp w(Xt ) − w(X0 ) − Gw(Xv ) dv .
0
For the case of standard Brownian motion B, take w(x) = µx. We shall soon learn that G = 21 ∆, and
1
Lt = exp(µBt − µ2 t)
2
is indeed a martingale.
Under the change of measure induced by L
Z
1 2 1 1 2
Q{Bt ∈ A} = E[IA (Bt ) exp(µBt − µ t)] = √ exp(µx − µ2 t)e−x /2t dx
2 2πt A 2
Z Z
1 1 2 2 2 1 1
= √ exp − (x − 2µxt − µ t ) dx = √ exp − (x − µt)2 dx
2πt A 2t 2πt A 2t
Z
1 2
= √ e−x /2t dx = P {Bt + µt ∈ A}
2πt A−µt
Exercise 4.100. Let B, P and Q be defined as above. For times 0 ≤ t1 < t2 · · · tn , and Borel sets A1 , . . . , An ,
Thus, by the Daniell-Kolmogorov extension theorem, under the measure Q, B is Brownian motion with drift
µ.
Example 4.101. Recall that for standard Brownian motion, and for a ≥ 0, the hitting time for level a,
τa = inf{t > 0; Bt > a}, has density
a 2
fτa (s) = √ e−a /2s .
2πt3
If we consider the process Bt +µt, µ > 0 under the measure Q, then this process is standard Brownian motion
under the measure P . Thus,
Q{τa ≤ t} = E[I{τa ≤t} Lt ] = E[E[I{τa ≤t} Lt |Fmin{τa ,t} ]] = E[I{τa ≤t} Lmin{τa ,t} ] = E[I{τa ≤t} Lτ ]
Z t
1 2 1 a 2
= E[I{τa ≤t} exp(µa − µ τa )] = exp(µa − µ2 s) √ e−a /2s ds
2 0 2 2πs 3
Thus,
a 2
fτa ,µ (s) = √ e−(a−µt) /2s
.
2πs3
gives the density of τa for Brownian motion with constant drift.
4 MARKOV PROCESSES 71
are independent of t ≥ 0.
Definition 4.103. A probabiity measure µ for a Markov process X is called stationary if X is a stationary
process under Pµ .
Lemma 4.104. µ is a stationary measure for a Markov process X if and only if
for all t ≥ 0.
Proof. The necessity is immediate. To prove sufficency, note that for each t ≥ 0, θt X is a Markov process
with initial distribution µ and the same generator as X, thus has the same finite dimensional distributions
as X.
Theorem 4.105. Let G generate a strongly continuous contraction semigroup T on C0 (S) corresponding to
a Markov process X. In addition, assume that D is a core for G. Then for a probability measure µ on S the
following are equivalent:
1. µ is a stationary distribution for X.
R R
2. T (t)f dµ = f dµ for f ∈ C0 (S) and t ≥ 0.
R
3. Gf dµ = 0 for f ∈ D.
Proof. (1 → 2) Z Z
T (t)f dµ = Eµ [T (t)f (X0 )] = Eµ [f (Xt )] = Eµ [f (X0 )] = f dµ
(2 → 1) By the lemma above, it suffices to show that the one dimensional distributions agree.
Z Z
Eµ [f (Xt )] = Eµ [T (t)f (X0 )] = T (t)f dµ = f dµ = Eµ [f (X0 )].
Rt R
2. If f dν = limn→∞ t1n 0 n T (t)f dα exists for every f ∈ C0 (S), and some increasing sequence
R
In other words, T (t) is a self adjoint operator with respect to the measure ν.
f0 = 1, f1 , . . . , fn , fn+1 = 1 ∈ Cb (S)
Set Z
I` = g` (x)f` (x)h` (x) ν(dx).
Then
n
Y n
Y
I0 = Eν [ fk (Xtk −t0 )], and In+1 = Eν [ fk (Xtn+1 −tk )].
k=1 k=1
Claim. I0 = In+1 .
Let 1 ≤ ` ≤ n, Then
Z Z
I` = g` f` h` dν = (g` f` )T (t`+1 − t` )(f`+1 h`+1 ) dν
Z
= T (t`+1 − t` )(g` f` )(f`+1 h`+1 ) dν
Z
= g`+1 f`+1 h`+1 dν = I`+1
because
n+1
!
Y
T (t`+1 − t` )(f`+1 (x)h`+1 (x)) = T (t`+1 − t` ) f`+1 (x)Ex [ fk (Xtk −t`+1 )]
k=`+2
n+1
Y n+1
Y
= T (t`+1 − t` )Ex [ fk (Xtk −t`+1 )] = Ex [ fk (Xtk −t` )] = h` (x),
k=`+1 k=`+1
and
`−1
!
Y
T (t`+1 − t` )(g` (x)f` (x)) = T (t`+1 − t` ) Ex [ fk (Xt` −tk )]f` (x)
k=1
`
Y
= T (t`+1 − t` )Ex [ fk (Xt` −tk )] = g`+1 (x).
k=1
and 4 follows.
R R
For a countable state space, the identity f1 Gf2 dν = f2 Gf1 dν becomes
XX XX
f1 (x)gxy (f2 (y) − f2 (x))ν{x} = f2 (x)gxy (f1 (y) − f1 (x))ν{x}.
x y x y
4 MARKOV PROCESSES 74
or by changing indices
XX XX
f1 (x)gxy f2 (y)ν{x} = f2 (y)gyx f1 (x)ν{y}.
x y x y
By writing f1 and f2 as sums of indicator functions of sites in S, we have that this statement is equivalent
to
ν{x}gxy = ν{y}gyx .
This is the equation of detailed balance.
Exercise 4.111. Find a reversible measure for a birth and death process and give conditions so that this
measure can be normalized to be a probability measure. Simplify this expression for the examples of birth and
death processes in the examples above.
Definition 4.112. A diffusion on I is a Feller process with paths in CI [0, ∞). A diffusion is called regular
if, for every x ∈ int(I) and y ∈ I,
Px {τy < ∞} > 0,
where τy = inf{t > 0 : Xt = y}.
Example 4.113. Brownian motion is a one dimensional diffusion on R. The semigroup is
1
Z
(x − y)2 1
Z √ y2
T (t)f (x) = √ f (y) exp(− ) dy = √ f (x + y t) exp(− ) dy.
2πt R 2t 2π R 2
1 1
Z
1 √ y2
(T (t)f (x) − f (x)) = √ ((f (x + y t) − f (x)) exp(− ) dy
t 2π R t 2
1
Z
1 √ 1 √ y2
= √ (y tf 0 (x) + y 2 tf 00 (x + θy t) exp(− ) dy
2π R t 2 2
1
Z
1 2 00 √ y 2
= √ y f (x + θy t) exp(− ) dy
2π R 2 2
Exercise 4.115. Let φ ∈ C 2 (R) be increasing and let B be standard Brownian motion. Find the generator
of X = φ ◦ B.
The major simplifying circumstance for one-dimensional diffusions is the intemediate value theorem, i.e.,
if x0 is between Xs and Xt , s < t, then Xc = x0 for some c ∈ (s, t). Regularity implies a much stronger
condition.
Theorem 4.116. Let J be a finite open interval, J¯ ⊂ I and let τJ be the time of the first exit from J, then
if X is a regular diffusion on I
sup Ex τJ < ∞.
x∈J
To complete the proof, set J = (a, b) and pick y ∈ J. Then, by the regularity of X, there exists a t > 0
and p < 1 so that
Py {τa > t} ≤ p, Py {τb > t} ≤ p.
Thus, for x ∈ [y, b),
Px {τJ > t} ≤ Px {τb > t} ≤ Py {τb > t}.
For x ∈ (a, y],
Px {τJ > t} ≤ Px {τa > t} ≤ Py {τa > t}.
To set notation, for any open interval J = (a, b) with Px {τJ < ∞} = 1 for all x ∈ J, write
Clearly, sJ (a) = 1 and sJ (b) = 0. If a < x < y < b, then by the strong Markov property,
Px {τa < τb } = Px {τd < τc }Pd {τa < τb } + Px {τc < τd }Pc {τa < τb }
Now, we extend. Take {Jn ; n ≥ 1} to be an increasing sequence of compact intervals with union I.
Moreover, if I contains an endpoint, then construct each Jn = [an , bn ] to include this endpoint. Define
s̃n : Jn → R inductively. s̃1 = sJ1 , and, for n ≥ 1,
Clearly, s̃n+1 (an ) = s̃n (an ), and s̃n+1 (bn ) = s̃n (bn ). Repeat the argument above to show that s̃n+1 is
a scale function on Jn . Because s̃n is an affine transformation of sn+1 and the two functions agree at two
points, they must be equal on Jn .
Consequently, we can define, for x ∈ Jn , s = s̃n .
Theorem 4.121. If X is a regular diffusion on I, then X̃ = s(X) is a regular diffusion in natural scale on
s(I).
Proof. Write ã = s(a), b̃ = s(b), and τ̃a = inf{t > 0 : X̃t = a}
s(x) − s(a) x̃ − ã
Px̃ {τ̃ã < τ̃b̃ } = Px {τa < τb } = = .
s(b) − s(a) b̃ − ã
Exercise 4.122. For a simple symmetric random walk Y on Z, and let τ = min{n; Yn ∈ {M− , M+ }},
M− < M+ and let Ny be the number of visits to y before τ . Then
Ex τ = (x − M− )(M+ − x),
and
(x−M− )(M+ −y)
(
M+ −M− , x≤y
Ex N y = (y−M− )(M+ −x)
M+ −M− , x ≥ y.
Theorem 4.123. Let X be a one dimensional diffusion on I on natural scale. Then there is a unique
measure m defined on B(int(I)) such that
4 MARKOV PROCESSES 78
Proof. Let {a = x0 < x1 < · · · < xn = b} be a partition [a, b] into subintervals of length h = (b − a)/n.
Thus, xk = a + hk Define Jk = (xk−1 , xk+1 ).
The process restricted to the exit times of the intervals Jk is a simple symmetric random walk Y on
(b − a) + hZ. Let σj denote the time between steps before τj and zero after. Then for a partition point x,
∞
X
Ex τ J = Ex σ j .
j=0
and
n−1
X
Ex σj = Ex [Ex [σj |Yj ]] = Px {Yj = xk }m(xk , Jk ).
k=1
Let n(xk0 , xk ) denote the expected number of visits of the random walk to xk starting at xk0 . Then, by the
exercise, (
k0 (n−k)
n xk0 ≤ xk 1
n(xk0 , xk ) = k(n−k0 ) = G(xk0 , xk ).
n xk ≥ xk0 h
Thus,
∞ n−1
X X n−1
X X∞
Ex τJ = Px {Yj = xk }m(xk , Jk ) = ( Px {Yj = xk })m(xk , Jk )
j=0 k=1 k=1 j=0
n−1 n−1
X X 1
= n(x, xk )m(xk , Jk ) = GJ (x, xk )m(x, Jk )
h
k=1 k=1
Define the measure mn that gives mass m(xk , Jk )/h to the point xk . Then the formula above becomes,
Z
m(xj , J) = GJ (xj , y) mn (dy).
J
Now consider succesive refinements, (say n = 2` ). Then the representation above holds whenever the
ratio (x − a)/(b − a) is a dyadic rational. This represenation above also implies that lim supn→∞ mn (J) < ∞.
Now use a variant of the Prohorov theorem to show that {m2` : ` ≥ 1} has compact closure in the topology
determined by weak convergence. Let m be a limit point for the {m2` : ` ≥ 1}, then for any finite open
interval J with dyadic rational endpoints, J¯ ⊂ I,
x−a
Z
m(x, J) = GJ (x, y) m(dy), a dyadic rational.
J b−a
By the bounded convergence theorem, this can be extended to all such open intervals J.
4 MARKOV PROCESSES 79
To extend this to arbitrary y, use the strong Markov property to obtain the identity for (ã, b̃) = J˜ ⊂ J,
m(x, J) = m(x, J˜n ) + Px {τã < τb̃ }m(b̃, J) + Px {τã > τb̃ }m(ã, J).
Now let ãn and b̃n be dyadic rationals ãn < x < b̃n , then
with ( (s(x)−s(a))(s(b)−s(y))
s(b)−s(a)) x≤y
GJ (x, y) = (s(y)−s(a))(s(b)−s(x))
s(b)−s(a)) x≤y
Ex τJ = (x − a)(x − b).
Ex τJ = (x − a)(x − b).
4. Use the standard machine to prove that for a bounded measurable function,
Z τJ Z
Ex [ f (Xt ) dt] = f (y)GJ (x, y) m(dy)
0 J
Theorem 4.127. Let X be a regular diffusion on natural scale with speed measure m and let X̃ be the unique
solution to
X̃t = XR t c(X̃s ) ds .
0
Proof. Set Z t
σt = c(X̃s ) ds.
0
Then, as before Z σt
t= c(X̃s ) ds.
0
Denote
τa = inf{t > 0 : Xt = a} τ̃a = inf{t > 0 : X̃t = a}.
Then
Px {τ̃a < τ̃b } = Px {τa < τb }.
Thus, X̃ is on natural scale.
Let τ̃J be the first exit time of X̃ from J. Then
Z τJ
1
τJ = στ̃J or τ̃J = ds.
0 c(Xs )
means that for any system of open sets {Jn ; n ≥ 1} decreasing to the point x, we have
lim φ(Jn ) = L.
n→∞
Ex f (XτJ ) − f (x)
Cf (x) = lim .
J→{x} E x τJ
The domain of C is the subspace of all f ∈ C0 (I) for which the limit above exists.
Theorem 4.130. Let X be a regular diffusion on I with generator A. Let f ∈ D(A) ∩ D(C), then Af (x) =
Cf (x) for all x ∈ int(I).
4 MARKOV PROCESSES 81
Proof. Choose f ∈ D(A), and x ∈ int(I). Set g = Af . Then Dynkin’s formula states that
Z τJ
Ex [f (XτJ )] = f (x) + E[ g(Xs ) ds].
0
for an interval J, J¯ ⊂ int(I). Choose > 0 so that g(J) ⊂ (g(x) − , g(x) + ). Then,
Ex f (XτJ ) − f (x)
g(x) − < < g(x) +
E x τJ
For f ∈ C 2 (R),
Ef (XτJ(λ,) ) = (1 − λ)f (x + λ) + λf (x − (1 − λ)).
and
Ef (XτJ(λ,) ) − f (x) = (1 − λ)(f (x + λ) − f (x)) + λ(f (x − (1 − λ)) − f (x)).
Now, choose sequences n → 0 and λn ∈ (0, 1). Then, by L’Hôpital’s rule
Set aside for the moment behavior in the boundary of I. Beginning with any scale s and any speed
measure having a density c, we can transform it to Brownian by a change of scale and a time change. Thus,
by inverting these transformations,
R we can transform a Brownian motion into a regular diffusion having scale
and speed measure A c(x) dx. This shows, among this class of scales and speed measure, that the behavior
of a regular diffusion is uniquely determined in the interior of I by these two properties of the diffusion.
For diffusions X in this class, the generator takes the form
1 d2 d
A= a(x) 2 + b(x) , a(x) > 0.
2 dx dx
1 d2
a(x)s0 (x)2 2 .
2 dx
Thus, the speed measure has density 1/(a(x)s0 (x)2 ).
Example 4.132. 1. The radial part of an n-dimensional Brownian motion R is called a Bessel process.
In this case, I = [0, ∞) and
1 d2 n−1 d
A= + .
2 dr2 2r dr
2. The Ornstein-Uhlenbeck process is obtained by subjecting particles undergoing physical Brownian mo-
tion to an elastic force. The generator is
σ 2 d2 d
A= − ρx , I = R.
2 dx2 dx
3. Financial models of stock prices use geometric Brownian motion. Here, I = (0, ∞) and the generator
is
σ 2 d2 d
A= x + ρx .
2 dx2 dx
4. A diffusion approximation of a neutral 2 allele model with scaled mutation rates µ0 and µ1 is
1 d2 d
A= x(1 − x) 2 + (µ1 x − µ0 (1 − x)) , I = [0, 1].
2 dx dx
Exercise 4.133. 1. Find a scale function and the speed measure for the four examples above.
2. Show that a normal distribution is stationary for the Ornstein-Uhlenbeck process.
3. Show that a beta distribution is stationary for the diffusion in example 4.
Definition 4.134. Let I = [r− , r+ ] ∩ R. The r+ is called an accessible is there exists t > 0 and x ∈ int(I)
such that
inf Px {τy ≤ t} > 0.
{y∈(x,r+ )}
4 MARKOV PROCESSES 83
If r+ is inaccessible, call it natural if there exists t > 0 and y ∈ int(I) such that
inf Px {τy ≤ t} > 0.
{y∈(x,r+ )}
An accessible boundary may or may not be regular. If it is not, then call the boundary point an exit.
Corresponding statements hold for r−
In schematic form, we have
r+ → int(I) r+ 6→ int(I)
int(I) → r+ regular exit
int(I) 6→ r+ entrance natural
Note that,
Ex τx0 is nondecreasing as x incresing.
and that
Z y Z x Z x
2(s(r+ ) − s(x) 2(s(c) − s(x0 )
G(x0 ,y) (x, s) m(dz) = (s(z)−s(x0 )) m(dz)+ (s(r0 )−s(z)) m(dz).
x0 2(s(r+ ) − s(x0 ) x0 2(s(r+ ) − s(x0 ) x0
This tends to zero as x → r+ . Thus,
lim Ex τx0 = 0
x→r+
and r+ is accessible.
5 STOCHASTIC INTEGRALS 84
5 Stochastic Integrals
In this section we develop the theory of stochastic integrals with the primary intent to apply them to the
study of a class of Markov processes having continuous sample paths.
The theory of Riemann-Steitjes integral applies readily to continuous functions having finite variation on
compact intervals. Thus, we can readily obtain a stochastic integral for
Z t
(X · V )t (ω) = Xs (ω) dVs (ω).
0
for
1. progressively measurable continuous V having finite variation on compact intervals, and
2. progressively measurable continuous X bounded on compact intervals.
To extend this to continuous martingales, we must take into account the following.
Exercise 5.1. Let M be a martingale with EMt2 < ∞ for all t > 0. Then, for s, t > 0,
2
Var(Ms+t |Ft ) = E[Ms+t |Ft ] − Mt2 = E[(Ms+t − Mt )2 |Ft ].
Thus
E[(Ms+t − Mt )2 ] = E[Ms+t
2
− Mt2 ].
Definition 5.2. A partition Π of [0, T ] is a strictly increasing sequence of real numbers 0 = t0 < t1 < · · · <
tn = T . A partition Π of [0, ∞) is a strictly increasing unbounded sequence of real numbers 0 = t0 < t1 <
· · · < tk · · · , limk→∞ tk = ∞. The mesh(Π) = sup{tk+1 − tk ; k ≥ 0}
For any process X and partition Π = {t0 < t1 < · · · < tk · · · } of [0, ∞), define the simple process
∞
X
XtΠ = Xtk I[tk ,tk+1 ) (t).
k=0
Theorem 5.3. A continuous martingale having finite variation on compact intervals is constant.
Proof. By considering the martingale M̃t = Mt − M0 , we can assume that M0 = 0. In addition, let σc be
the time that the variation of M reaches C. Then by considering the martingale M σc , we can assume that
M has variation bounded above by c.
Now, let Π = {0 = t0 < t1 < · · · < tn−1 < tn = t}, be a partitions of [0, t], then
n
X n
X
EMt2 = E[Mt2j+1 − Mj2 ] = E[(Mtj+1 − Mtj )2 ] ≤ nE[ max |Mtj+1 − Mtj |].
1≤j≤k
j=1 j=1
The random variable max1≤j≤k |Mtj+1 − Mtj | is bounded above by c and converges to 0 almost surely as the
mesh of the partition tends to 0. Thus,
EMt2 = 0 and Mt = 0 a.s.
5 STOCHASTIC INTEGRALS 85
We say that X has finite quadratic variation if there exists a process hX, Xi such that
QΠ P
t (X) → hX, Xit as mesh(Π) → 0.
Theorem 5.5. A continuous bounded martingale M has finite quadratic variation hM, M it that is the unique
continuous increasing adapted process vanishing at zero such that
Mt2 − hM, M it
is a martingale.
Proof. The uniquess of hM, M i is easy, let A and B be two such processes. Then,
At − Bt = (Mt2 − Bt ) − (Mt2 − At )
as the difference of two martingales is itself a martingale. Because it has finite variation on compact intervals,
it is constant.
For existence, we will prove:
Exercise 5.7. Let B denote standard one-dimensional Brownian motion. With the partitions Πn described
L2
as above, show that QΠ
t (B) →
n
t as n → ∞.
n
For this, we will introduce a second
R t sequence I of process, which we shall later recognize as the approx-
imation to the stochastic integral 0 Ms dMs . Write Πn = {0 = tn0 < tn1 < · · · < tnk · · · }, then
∞
X
ItΠn = Mmin{t,tnk } (Mmin{t,tnk+1 } − Mmin{t,tnk } ).
k=0
Exercise 5.8. 1.
Mt2 − M02 = QΠ Πn
t (M ) + 2It .
n
(5.1)
E[(ItΠn )2 ] ≤ c4 .
Thus,
E[QΠ n 2 Πn 2 2 2 4
t (I )] ≤ c E[Qt (M )] = c E[Mt − M0 ] ≤ c .
n
Proof. Write
Πn ∪ Πm = {u0 < u1 < · · · < uk · · · }.
Then
∞
X
ItΠn − ItΠm = Πn
(Mmin{t,u k}
Πm
− Mmin{t,u k}
)(Mmin{t,uk+1 } − Mmin{t,uk } )
k=1
m ∪Πm
is a mean zero martingale, and therefore, (ItΠn − ItΠm )2 − QΠ
t (I Πn − I Πm ) is, by the exercise, a mean
zero martingale.. Thus,
∞
X
m ∪Πm
QΠ
t (I Πn − I Πm ) = Πn
(Mmin{t,u k}
Πm
− Mmin{t,u k}
)2 (Mmin{t,uk+1 } − Mmin{t,uk } )2
k=1
m ∪Πm
≤ sup (MsΠn − MsΠm )2 QΠ
t (M ).
0≤s≤t
In addition,
m ∪Πm
E[QΠ
T (M )2 ] ≤ 2(E[(MT2 − M02 )2 ] + 4E[(ITΠm ∪Πn )2 ]) ≤ 12c4 .
Finally, sup0≤t≤T (MtΠn − MtΠm )4 is bounded and converges to zero almost surely as m, n → ∞. Thus, by
the bounded converges theorem, its expectation converges to zero and (5.2) holds.
Returning to the proof of the theorem.
5 STOCHASTIC INTEGRALS 87
Proof. The continuous processes I Πn are Cauchy in the norm ||X − Y ||2,T = E[sup0≤t≤T (Xt − Yt )2 ]1/2 , thus,
by (5.1). so is QΠn (M ). Because the space is complete and the convergences is uniform almost surely, the
limit Q(M ) is continuous. Choose, if necessary, a subsequence that converges almost surely.
Write Π̃n = ∪nk=1 Πn . Then for any s, t ∈ Π̃n , s < t
QΠ̃ Π̃n
t (M ) ≥ Qs (M ).
n
Consequently, Q(M ) is nondecreasing on the dense set ∪∞ k=1 Πn . Because Q(M ) is continuous, it is nonde-
creasing. Finally, that fact that Mt2 −Qt (M ) is a martingale can be seen by taking limits on the expectations
hM τ , M τ it = hM, M iτt .
(M τ )2 − hM, M iτ .
Mt2 − hM, M it
is a local martingale.
In addition, for any T > 0 and any sequence of partitions {Πn ; n ≥ 1} of [0, ∞) with mesh(Πn ) → 0 then
sup |QΠ P
t (M ) − hM, M it | → 0
n
0≤t≤T
as n → ∞.
Proof. Choose {τm ; m ≥ 1} be a reducing sequence for M so that M τm I{τm >0} is bounded. Let Qm (M )
denote the unique continuously increasing process, Qm
0 (M ) = 0 so that
(M τm )2 I{τm >0} − Qm (M )
(Qm+1 (M ))τm = Qm (M )
QΠ
t (M
n τm0
) = QΠ
t (M )
n
and hM τm0 , M τm0 it = hM, M it .
Consequently,
P { sup |QΠ Πn
t (M ) − hM, M it | > } ≤ δ + P { sup |Qt (M
n τm0
) − hM τm0 , M τm0 it | > }.
0≤t≤T 0≤t≤T
Theorem 5.13. Let M and N be two continuous local martingales, then there is a unique continuous adapted
process hM, N i of bounded variation vanishing at zero so that
Mt Nt − hM, N it
is a local martingale.
In addition, set, for the partition Π for 0 = t0 < t1 < · · · < tk · · · , limk→∞ tk = ∞.
∞
X
CtΠ (M, N ) = (Mmin{t,tj+1 } − Mmin{t,tj } )(Nmin{t,tj+1 } − Nmin{t,tj } ).
j=1
Then, for any t > 0 and any sequence of partitions {Πn ; n ≥ 1} of [0, ∞) with mesh(Πn ) → 0 then
as n → ∞.
Proof. Set
1
hM, N it =(hM + N, M + N it ) − hM − N, M − N it )
4
and use an appropriate reducing sequence. Uniqueness follows are before.
Definition 5.14. We will call hM, N i the bracket of M and N and hM, M i the increasing process associated
with M
Soon, we shall develop the stochastic integral. We will begin with a class of simple functions and extend
using a Hilbert space isometry. The Kunita-Watanabe inequalities are central to this extension.
Proposition 5.17. Let M and N be local martingales and let H and K be measurable processes, then
Z t Z t 1/2 Z t 1/2
| Hs Ks dhM, N is | ≤ Hs2 dhM, M is Ks2 dhN, N is .
0 0 0
We now introduce the class of processes that will become the stochastic integraters.
5 STOCHASTIC INTEGRALS 90
X =M +V (5.3)
where M is a continuous local martingale and V is a continuous adapted process having finite variation on
a sequence of stopping time {τn ; n ≥ 1}, limn→∞ τn = ∞ bounded time intervals.
Exercise 5.20. 1. Show that V has zero quadratic variation.
2. Using the notation above show that hM, V i = limmesh(Πn )→0 C Πn (M, V ) = 0.
hX, Y i = hM, N i.
Definition 5.21. 1. Define H2 to be the space of L2 -bounded martingales. These are martingales M ,
such that
sup EMt2 < ∞.
t≥0
Use H02 2
for those elements M ∈ H with M0 = 0.
2. For M ∈ H2 define L2 (M ) to be the space of progressively measurable processes K such that
Z ∞
||K||2M = E[K 2 · hM, M i] = E[ Ks2 dhM, M is ] < ∞.
0
If M ∈ H2 , then M is a uniformly integrable martingale, the its limit exists and is in L2 (P ). Thus we
can place the following norm
||M ||2H2 = EM∞
2
.
Proposition 5.22. A continuous local martingale M is in H02 if and only if EhM, M i∞ < ∞. In this case,
M 2 − hM, M i is a uniformly integrable martingale.
Proof. Let {τn ; n ≥ 1} be a sequence that reduces M . Then,
2
E[Mmin{t,τ I
n } {τn >0}
] − E[hM, M imin{t,τn } I{τn >0} ] = E[M02 I{τn >0} ].
∗
Because M∞ ∈ L2 (P ), we can use the dominated and monotone convergence theorem to conclude that
2
E[M∞ ] − E[hM, M i∞ ] = E[M02 ]
5 STOCHASTIC INTEGRALS 91
By Fatou’s lemma,
EMt2 ≤ lim inf E[Mmin{t,τ
2
I
n } {τn >0}
]≤C
n→∞
Thus, {Mmin{t,τn } I{τn >0} ; n ≥ 1, t ≥ 0} is uniformly integrable. Take s < t and A ∈ Fs , then
E[Mt ; A] = lim E[Mmin{t,τn } I{τn >0} ; A] = lim E[Mmin{s,τn } I{τn >0} ; A] = E[Ms ; A]
n→∞ n→∞
and M is a martingale.
Finally, note that
sup |Mt2 − hM, M it | ≤ (M 2 )∗∞ + hM, M i∞ ,
t≥0
Note that
X n
n X
E(K · M )2 = E[ Kj (Mtj − Mtj−1 )Kk (Mtk − Mtk−1 )].
j=1 k=1
If j < k,
E[Kj (Mtj − Mtj−1 )Kk (Mtk − Mtk−1 )] = E[Kj (Mtj − Mtj−1 )Kk E[Mtk − Mtk−1 |Ftk−1 ]] = 0.
Thus,
Xn Xn
E(K · M )2 = E[ Kj2 (Mtj − Mtj−1 )2 ] = E[ Kj2 E[(Mtj − Mtj−1 )2 |Ftj−1 ]]
j=1 j=1
Xn Xn
= E[ Kj2 E[(Mt2j − Mt2j−1 |Ftj−1 ]] = E[ Kj2 E[hM, M itj − hM, M itj−1 |Ftj−1 ]]
j=1 j=1
Xn Z ∞
= E[ Kj2 (hM, M itj − hM, M itj−1 )] = E[ Ks2 dhM, M is ] = ||K||2M
j=1 0
5 STOCHASTIC INTEGRALS 92
Exercise 5.23. Let N and M be square integrable martingales and let H and K be simple adapted functions,
then
E[(K · M )(H · N )] = E[(KH · hM, N i)].
Theorem 5.24. Let M ∈ H02 , then for each K ∈ L2 (M ), there is a unique element K · M ∈ H02 , such that
hK · M, N i = K · hM, N i.
Proof. To verify uniqueness, note that if M 1 and M 2 are two martingales in H02 , such that
Then,
hM 1 − M 2 , M 1 − M 2 i = 0.
and M 1 − M 2 is constant, hence 0.
By the Kunita-Watanabe inequality,
Z ∞
|E[ Ks dhM, N is ]| ≤ ||N ||H2 ||K||L2 (M ) .
0
and thus
K →K ·M
is an isometry.
Definition 5.25. The martingale K · M is called the Itô stochastic integral of K with respect to M . It is
often also denoted by Z t
(K · M )t = Ks dMs .
0
(HK) · M = H · (K · M ).
Proof. Apply the theorem above twice to see that hK · M, K · M i = K 2 hM, M i. Thus, we have H · K ∈
L2 (K · M ). For N ∈ H02 , we also have
K · M τ = KI[0,τ ] · M = (K · M )τ .
We now we shall use localization to extend the definition of the stochastic integral.
Definition 5.28. Let M be a continuous local martingale, then L2loc (M ) is the space of progressively mea-
surable functions K for which there exists an increasing sequence of stopping times {τn ; n ≥ 1} increasing
to infinity so that Z τn
E[ Ks2 dhM, M is ] < ∞.
0
Exercise 5.29. For any K ∈ L2loc (M ), there exists a continuous local martingale K · M such that for any
continuous local martingale N ,
hK · M, N i = K · hM, N i.
Definition 5.30. 1. Call a progressively measurable process K locally bounded if there exists a sequence
of stopping times {τn ; n ≥ 1} increasing to infinity and constants Cn such that K τn ≤ Cn .
2. Let K be locally bounded and let X = M +V be a continuous semimartingale, X0 = 0, the Itô stochastic
integral of K with respect to X is the continuous semimartingle
K ·X =K ·M +K ·V
where K · M is the stochastic integral defined above and K · V is the usual Stieltjes integral.
Exercise 5.31. The map K → K · X has the following properties:
1. H · (M · X) = (HM ) · X.
2. (K · M )τ = (KI[0,τ ] ) · M ) = K · M τ .
3. If X is a local martingale or a process of finite variation, so is K · X.
4. If K is an progressively measurable simple function,
∞
X
K = K0 I{0} + Kj I(tj−1 ,tj ] , 0 = t0 < t1 < · · · , lim tj = ∞,
j→∞
j=1
then
∞
X
(K · X)t = Kj (Xmin{t,tj } − Xmin{t,tj−1 } .
j=1
Theorem 5.32 (bounded convergence theorem). Let X be a continuous semimartingale and let {K n ; n ≥ 1}
be locally bounded processes converging to zero pointwise with |K n | ≤ K for some locally bounded process K,
then for any T > 0 and > 0,
lim P { sup |(K n · X)t | > } = 0.
n→∞ 0≤t≤T
Proof. If X is a process of finite variation, then we can apply the bounded convergence theorem for each
ω ∈ Ω.
If X is a local martingale and it τ reduces X, then by the bounded convergence theorem, (K n )τ converges
to zero in L2 (X τ ) and thus (K n · X)τ converges to zero in H2 . Now repeat the argument on the convergence
of the quadratic variation over a partition for a local martingale.
5 STOCHASTIC INTEGRALS 95
Exercise 5.33. Let K be left-continuous and let {Πn ; n ≥ 1} be a sequence of partitions of [0, ∞), with
Πn = {0 = tn0 < tn0 < · · · }, mesh(Πn ) → 0, then
∞
X
lim P { sup |(K · X)t − Ktnj−1 (Xmin{t,tnj } − Xmin{t,tnj−1 } | > } = 0.
n→∞ 0≤t≤T j=1
Exercise 5.34. 1. Let X be a continuous semimartingale and let the stochastic integrals H · X and H · X
exists, then for a, b ∈ R
Z t Z t Z t
(aHs + bKs ) dXs = a Hs dXs + b Ks dXs .
0 0 0
2. Let X and X̃ be two continuous semimartingales and let K and K̃ be two predictable locally bounded
processes. Prove that if the finite dimensional distributions of (K, X) and (K̃, X̃) agree, then so do the
finite dimensional distributions of (K, X, K · X) and (K̃, X̃, K̃ · X̃).
Proof. Let {Πn ; n ≥ 1} be a sequence of partitions of [0, ∞), Πn = {0 = tn0 < tn0 < · · · }, mesh(Πn ) → 0.
Expand to obtain,
∞
X ∞
X
(Xmin{t,tj } − Xmin{t,tj−1 } )2 = Xt2 − X02 − 2 Xmin{t,tj−1 } (Xmin{t,tj } − Xmin{t,tj−1 } ).
j=1 j=1
Then,
∞
X
(Xmin{t,tj } − Xmin{t,tj−1 } )2 → hX, Xit .
j=1
and
∞
X Z t
Xmin{t,tj−1 } (Xmin{t,tj } − Xmin{t,tj−1 } ) → Xs dXs
j=1 0
Proof. The corollary follows from polarization of the identity in the previous proposition.
Example 5.37. Let B be standard Brownian motion, then
Z t
2
Bt − t = Bs dBs .
0
Definition 5.38. A d-dimensional vector (continuous) local martingale is an Rd -valued stochastic process
X = (X 1 , . . . , X d ) such that each component is a (continuous) local martingale. A complex (continuous)
local martingale is one in which both the real and imaginary parts are (continuous) local martingales.
Theorem 5.39 (Itô formula). Let f ∈ C 2 (Rd , R) and X = (X 1 , . . . , X d ) be a vector continuous semimartin-
gale, then f ◦ X is a continuous semimartingale and
d Z t d d t
∂2f
Z
X ∂f 1 XX
f (Xt ) = f (X0 ) + (Xs ) dXsi + (Xs ) dhX i , X j is .
i=1 0 ∂xi 2 i=1 j=1 0 ∂xi ∂xj
in differential form
d
X
dYt = Hsi dXsi .
i=1
λ2
Et (λM ) = exp(λMt − hM, M it ), λ∈C
2
5 STOCHASTIC INTEGRALS 97
is a local martingale.
This local martingale satisfies the stochastic differential equation
λ2
exp(λBt − t)
2
is a martingale. Use the standard machine to show that
Z t
1 t
Z
Etf (B) = exp f (t) dBs − f (t)2 ds
0 2 0
is a martingale.
3. If B is a d-dimensional Brownian motion and if f ∈ C 2 (R+ × Rd ), then
Z t
1 ∂f
f (t, Bt ) − ( ∆f + )(s, Bs ) ds
0 2 ∂t
Theorem 5.43 (Lévy’s characterization theorem). Assume that X is a Rd valued stochastic process satis-
fying X0 = 0. Then the following are equivalent.
1. X is d-dimensional standard Brownian motion.
2. X is a continuous local martingale and hX i , X j i = δij t.
hM, M it = |ξ|2 t
We will be focused on pathwise uniquesness. We begin with the basic estimate used to study the
dependence of a solution on initial condiations.
Theorem 5.46. (Gronwall’s inequality) Let f, g : [0, T ] → R be continuous and nonnegative. Suppose
Z t
f (t) ≤ A + f (s)g(s) ds, A ≥ 0.
0
Theorem 5.47. Let Ft be a right-continuous complete filtration on (Ω, FP ) and let Z be a continuous
r-dimensional semimartingale. If the progressive process f (t, X) satisfies the the Lipschitz condition above
and if for every t > 0,
Z s 2
E[ sup f (r, x̄) dr ] < ∞,
0≤s≤t 0
r
where x̄ is a constant function for some x0 ∈ R , then there is a pathwise unique process X such that
Z t
Xt = x + f (s, X) dZs .
0
Proof. Note that the Lipschitz condition guarantees that if the condition above holds for one choice x0 then
it holds for all x ∈ Rr .
Write the semimartingale Z = M + A in its canonical decomposition. Let |A|t denote the variation of A
up to time t
5 STOCHASTIC INTEGRALS 100
Case 1. The measures associated to hM, M i and |A| are absolutely continuous with respect to Lebesgue
measure and for all i and j.
dhM i , M j it d|A|t
≤ 1, ≤ 1.
dt dt
X r Z t
+2tE[ (fj (r, U ) − fj (r, V ))2 d|Aj |s |]
j=1 0
Z t
≤ 2K 2 (4r2 + rt) ||U − V ||22,r dr
0
X 0 = x, and X n = SX n−1
and write
C = 2K 2 (4r2 + rT ).
Now, check, by induction, that
(Ct)n
||X n+1 − X n ||22,t ≤ ||X 1 − X 0 ||22,T .
n!
Thus, {X n ; n ≥ 1} is a Cauchy sequence in || · ||2,T . Call the limit X. Then X is a continuous process
and satisfies X = XS. In other words, X is a solution.
To show that X is unique, let X̃ be a second solution. Define τk = inf{t > 0 : |Xt | ≥ k or |X̃l | ≥ k}. Use
Gronwall’s inequality, with f (t) = ||X τk − X̃ τk ||2,t and g(t) = C to see that X̃0 = x, implies X̃ τk = X τk .
Now, let T and k go to infinity.
Case 2. The general case.
5 STOCHASTIC INTEGRALS 101
Then Z̃ satisfies the circumstances in case 1 with the filtration F̃t = FCt . Thus,
Z t
X̃t = x + f (s, X̃) dZ̃s
0
has a unique solution. Now, Xt = X̃Ãt is the unique solution to the equation in the statement of the
theorem.
Theorem 5.49. Let σ and b be locally bounded and Borel measurable, and let α be a probability measure on
Rd . Then pathwise uniqueness to the Itô stochastic differential equation above implies distribution uniqueness.
We begin with a lemma.
X0 = x, X̃0 = x, a.s.
Proof. Let P be the distribution of (X, B) on the space CRr [0, ∞) and write the pair (ξ, β) for the coordinate
functions for the first d and last r coordinates respectively. Because CRd+r [0, ∞) is a Polish space, there is
a regular conditional probability
P (B|B0 )(ξ, β) = P (B, (ξ, β)).
In particular, β has the distribution of r-dimensional Brownian motion. The triple (X, B, σ(X) · B) and
(ξ, β, σ(ξ) · β) have the same distribution and thus (ξ, β) solves the stochastic differential equation.
Repeat this letting P̃ be the distribution of (X̃, B̃). Then, the regular conditional probabilities agree
P̃ (B, (ξ, β)) = P (B, (ξ, β))
and so if X0 and X̃0 have the same distribution, then P = P̃ .
Theorem 5.51. If pathwise uniqueness holds for an Itô type stochastic differential equation, then uniqueness
in law holds. Moreover, there exists a measureable map
F : CRd [0, ∞) → CRr [0, ∞)
such that F (β) is adapted to the natural filtration Btr = σ{ξs ; s ≤ t} and for every solution (X, B), we have
that
Xt (ω) = F (B(ω))t a.s.
Proof. Let (X, B) be a solution on (Ω, P ) and let (X̃, B̃) be a solution on (Ω̃, P̃ ). With, for some x ∈ Rd ,
P {X0 = x} = P̃ {X̃0 = x} = 1.
For the solution (X, B) let Q be the image of P under the map
ω → (X(ω), B(ω)).
For the solution (X̃, B̃) let Q̃ be the image of P̃ under the map
ω̃ → (X̃(ω̃), B̃(ω̃)).
In addition, let W be Wiener measure, that is the law of r-dimensional Brownian motion, Define a probability
measure Π on CRd [0, ∞) × CRd [0, ∞) × CRd [0, ∞) by
˜ dβ) = Q(β, dξ)Q̃(β, dξ)W
Π(dξ, dξ, ˜ (dβ)
with
d d d
X ∂f 1 XX ∂2f
Gf (x) = bi (x) (x) + aij (x) (x)
i=1
∂xi 2 i=1 j=1 ∂xi ∂xj
T
and a = σσ .
square roots
Definition 5.53. A stochastic process X is called a solution to the martingale problem for the collection A
if Z t
f (Xt ) − g(Xs ) ds
0
Lemma 5.55. Assume that σ and b are locally bounded and measurable. In addition, σ −1 exists and is
locally bounded. If X is a solution to the CRd [0, ∞)-martingale problem for (A, α) with respect to a complete
filtration {Ft ; t ≥ 0}, then there exists a d-dimensional Brownian motion B such that (X, B) is a solution
to the stochastic differential equation.
Proof. For every f ∈ Cc∞ (Rd )
Z t
Mtf = f (Xt ) − f (X0 ) − Af (Xs ) ds
0
is a continuous martingale, we have that this expression gives a continuous local martingale for every f ∈
C ∞ (Rd ). Choose f (x) = xi to obtain the martingale
Z t
i i
Mt = X t − X 0 − bi (Xs ) ds
0
i j
is a local martingale. Choose f (x) = x x ,
Z t
hM i , M j it = hX i , X j it = aij (Xs ) ds.
0
Claim. Z t
Bt = σ −1 (Xs ) dMs
0
is a d-dimensional Ft -adapted Brownian motion.
B is a vector valued continuous local martingale and B0 = 0. Thus, by Lévy’s characterization, it suffices
to show that hB i , B j it = δij t.
d X
X d
hB i , B j it = hσ −1 (X)ik · M k , σ −1 (X)j` · M ` it
k=1 `=1
d X
X d Z t
−1
= (σik ak` (σ T )−1
`j )(Xs ) ds = δij t.
k=1 `=1 0
Thus, Z t Z t Z t
σ(Xs ) dBs = dMs = Xt − X0 b(Xs ) ds
0 0 0
and the stochastic differential equation has a solution.
5 STOCHASTIC INTEGRALS 105
To show that this Brownian can be used to solve the stochastic differential equation, note that,
Z t Z t Z t
σ(X̃s ) dB̃s = (σρ)(X̃s ) dM̃s + (ση)(X̃s ) dB̃s
0 0 0
Z t Z t
= (σρ)(X̃s ) dM̃s = M̃t − (Id − σρ)(X̃s ) dM̃s
0 0
Z t
= M̃t = X̃t X̃0 − b(Xs ) ds
0
because
d
X d
X d X
X d Z t
h (Id − σρ)(X̃s )ik · M̃ k , (Id − σρ)ik (X̃s ) · M̃ k it = (Id − σρ)ik ak` (Id − σρ)i` (X̃s ) ds = 0
k=1 `=1 k=1 `=1 0